Modeling Hippocampal and Neocortical Contributions to...

Modeling Hippocampal and Neocortical Contributions to RecognitionMemory: A Complementary-Learning-Systems Approach

Kenneth A. Norman and Randall C. O’ReillyUniversity of Colorado at Boulder

The authors present a computational neural-network model of how the hippocampus and medial temporallobe cortex (MTLC) contribute to recognition memory. The hippocampal component contributes byrecalling studied details. The MTLC component cannot support recall, but one can extract a scalarfamiliarity signal from MTLC that tracks how well a test item matches studied items. The authors presentsimulations that establish key differences in the operating characteristics of the hippocampal-recall andMTLC-familiarity signals and identify several manipulations (e.g., target–lure similarity, interference)that differentially affect the 2 signals. They also use the model to address the stochastic relationshipbetween recall and familiarity and the effects of partial versus complete hippocampal lesions onrecognition.

Memory can be subdivided according to functional categories(e.g., declarative vs. procedural memory; Cohen & Eichenbaum,1993; Squire, 1992b) and according to neural structures (e.g.,hippocampally dependent vs. nonhippocampally dependent formsof memory). Various attempts have been made to align thesefunctional and neural levels of analysis; for example, Squire(1992b) and others have argued that declarative memory dependson the medial temporal lobe whereas procedural memory dependson other cortical and subcortical structures. Recently, we and ourcolleagues have set forth a computationally explicit theory of howhippocampus and neocortex contribute to learning and memory(the complementary-learning-systems model; McClelland, Mc-Naughton, & O’Reilly, 1995; O’Reilly & Rudy, 2001). In thisarticle, we advance the complementary-learning-systems model byusing it to provide a comprehensive treatment of recognition-memory performance.

In this introductory section, we describe two questions that haveproved challenging for math-modeling and cognitive-neuroscienceapproaches to recognition, respectively: In the math-modelingliterature, there has been considerable controversy regarding howto characterize the contribution of recall (vs. familiarity) to recog-nition memory; in the cognitive-neuroscience literature, research-ers have debated how the hippocampus (vs. surrounding corticalregions) contributes to recognition. Then, we show how our mod-

eling approach, which is jointly constrained by behavioral andneuroscientific data, can help resolve these controversies.

Dual-Process Controversies

Recognition memory refers to the process of identifying stimulior situations as having been experienced before, for example,when one recognizes a person one knows in a crowd of strangers.Recognition can be compared with various forms of recall memorywhere specific content information is retrieved from memory andproduced as a response; recognition does not require recall ofspecific details (e.g., one can recognize a person as being familiarwithout being able to recall who exactly the person is or where oneknows the person from). Nevertheless, recognition can certainlybenefit from recall of specific information—if one can recall thata familiar person at the supermarket is in fact one’s veterinarian,that reinforces the feeling that one actually does know this person.Theories that posit that recognition is supported by specific recallas well as by nonspecific feelings of familiarity are called dual-process theories (see Yonelinas, 2002, for a thorough review ofthese theories).

Although it is obvious that recall can (in principle) contribute torecognition judgments, the notion that recall routinely contributesto item-recognition performance is quite controversial. Practicallyall extant math models of recognition consist of a unitary famil-iarity process that indexes in a holistic fashion the global matchbetween the test probe and all of the items stored in memory (see,e.g., Gillund & Shiffrin, 1984; Hintzman, 1988; Humphreys, Bain,& Pike, 1989). These familiarity-only models can explain a verywide range of recognition findings (for reviews, see Clark &Gronlund, 1996; Raaijmakers & Shiffrin, 1992; Ratcliff & Mc-Koon, 2000)—even findings that, at first glance, appear to requirea recall process (see, e.g., McClelland & Chappell, 1998). Further-more, the relatively small number of findings that cannot beexplained using standard global-matching models tend to comefrom specialized paradigms like Jacoby’s process-dissociation pro-cedure (Jacoby, 1991; see Ratcliff, Van Zandt, & McKoon, 1995,for discussion of when global-matching models can and cannot

Kenneth A. Norman and Randall C. O’Reilly, Department of Psychol-ogy, University of Colorado at Boulder.

This work was supported by Office of Naval Research Grant N00014-00-1-0246, National Science Foundation Grant IBN-9873492, and Na-tional Institutes of Health (NIH) Program Project MH47566. Kenneth A.Norman was supported by NIH National Research Service Award Fellow-ship MH12582. We thank Rafal Bogacz, Neil Burgess, Tim Curran, DavidHuber, Michael Hasselmo, Caren Rotello, and Craig Stark for their veryinsightful comments on a draft of this manuscript.

Correspondence concerning this article should be addressed to KennethA. Norman, who is now at the Department of Psychology, PrincetonUniversity, Green Hall, Princeton, New Jersey 08544. E-mail:[email protected]

Psychological Review Copyright 2003 by the American Psychological Association, Inc.2003, Vol. 110, No. 4, 611–646 0033-295X/03/$12.00 DOI: 10.1037/0033-295X.110.4.611

611

account for process-dissociation data). As such, it is always pos-sible to treat these findings as special cases that have little rele-vance to performance on standard item-recognition tests.

Another issue that has hindered the acceptance of dual-processmodels is the difficulty inherent in measuring the separate contri-butions of recall and familiarity. Several techniques have beendevised for quantitatively estimating how recall and familiarity arecontributing to recognition performance (receiver operating char-acteristic [ROC] analysis, independence remember–know, andprocess dissociation; see Yonelinas, 2001, 2002, for review anddiscussion), but all of these techniques rely on a core set ofcontroversial assumptions; for example, they all assume that recalland familiarity are stochastically independent. There are reasons tobelieve that the independence assumption may not always be valid(see, e.g., Curran & Hintzman, 1995). Furthermore, there is no wayto test this assumption using behavioral data alone because ofchicken-and-egg problems (i.e., one needs to measure familiarityto assess its independence from recall, but one needs to assumeindependence to measure familiarity).

These chicken-and-egg problems have led to a rift between mathmodelers and other memory researchers. On the empirical side,there is now a vast body of data on recall and familiarity, gatheredusing measurement techniques that assume (among other things)independence—these data could potentially be used to constraindual-process models. However, on the theoretical side, modelersare not making use of these data because of reasonable concernsabout the validity of the assumptions used to collect them andbecause single-process models have been quite successful at ex-plaining recognition data (so why bother with more complexdual-process models?). To resolve this impasse, one needs somesource of evidence that one can use to specify the properties ofrecall and familiarity other than the aforementioned measurementtechniques.

Cognitive-Neuroscience Approaches to RecognitionMemory

Just as controversies exist in the math-modeling literature re-garding the contribution of recall to recognition memory, parallelcontroversies exist in the cognitive-neuroscience literature regard-ing the contribution of the hippocampus to recognition memory.

Researchers have long known that the medial temporal region ofthe brain is important for recognition memory. Patients with me-dial temporal lobe lesions encompassing both the hippocampusand surrounding cortical regions (perirhinal, entorhinal, and para-hippocampal cortices, which we refer to jointly as medial temporallobe cortex [MTLC]) typically show impaired recall and recogni-tion but intact performance on other memory tests (e.g., perceptualpriming, skill learning; see Squire, 1992a, for a review).

The finding of impaired recall and recognition in medial tem-poral amnesics is the basis for several influential taxonomies ofmemory. Most prominently, Squire (1987, 1992b), Eichenbaumand Cohen (Cohen & Eichenbaum, 1993; Cohen, Poldrack, &Eichenbaum, 1997; Eichenbaum, 2000), and others have arguedthat the medial temporal lobes implement a declarative memorysystem, which supports recall and recognition, and that other brainstructures support procedural memory (e.g., perceptual priming,motor-skill learning). Researchers have argued that the medialtemporal region is important for declarative memory because it is

located at the top of the cortical hierarchy and therefore is ideallypositioned to associate aspects of the current episode that are beingprocessed in domain-specific cortical modules (see, e.g., Mishkin,Suzuki, Gadian, & Vargha-Khadem, 1997; Mishkin, Vargha-Khadem, & Gadian, 1998). See Figure 1 for a schematic diagramof how hippocampus, MTLC, and neocortex are connected.

Although the basic declarative-memory framework is widelyaccepted, attempts to tease apart the contributions of differentmedial temporal structures have been more controversial. There iswidespread agreement that the hippocampus is critical for recall—focal hippocampal lesions lead to severely impaired recall perfor-mance. However, the data are much less clear regarding effects offocal hippocampal damage on recognition. Some studies havefound roughly equal impairments in recall and recognition (see,e.g., Manns & Squire, 1999; Reed, Hamann, Stefanacci, & Squire,1997; Reed & Squire, 1997; Rempel-Clower, Zola, & Amaral,1996; Zola-Morgan, Morgan, Squire, & Amaral, 1986), whereasother studies have found relatively spared recognition after focalhippocampal lesions (see, e.g., Holdstock et al., 2002; Mayes,Holdstock, Isaac, Hunkin, & Roberts, 2002; Vargha-Khadem etal., 1997). The monkey literature parallels the human literature—some studies have found relatively intact recognition (indexedusing the delayed nonmatch-to-sample test) following focal hip-pocampal damage (see, e.g., Murray & Mishkin, 1998), whereasothers have found impaired recognition (see, e.g., Beason-Held,Rosene, Killiany, & Moss, 1999; Zola et al., 2000). Spared rec-ognition following hippocampal lesions depends critically onMTLC—whereas recognition is sometimes spared by focal hip-pocampal lesions, it is never spared after lesions that encompassboth MTLC and the hippocampus (see, e.g., Aggleton & Shaw,1996).

Aggleton and Brown (1999) have tried to frame the differencebetween hippocampal and cortical contributions in terms of dual-

Figure 1. Schematic box diagram of neocortex, medial temporal lobecortex (MTLC), and the hippocampus. MTLC serves as the interfacebetween neocortex and the hippocampus. MTLC is located at the very topof the cortical processing hierarchy—it receives highly processed outputsof domain-specific cortical modules, integrates these outputs, and passesthem on to the hippocampus; it also receives output from the hippocampusand passes this activation back to domain-specific cortical modules viafeedback connections. Adapted from “The Medial Temporal MemorySystem,” by L. R. Squire and S. Zola-Morgan, 1991, Science, 253, p. 1380.Copyright 1991 by the American Association for the Advancement ofScience (http://www.sciencemag.org). Adapted with permission.

612 NORMAN AND O’REILLY

process models of recognition. According to this view, (a) thehippocampus supports recall, and (b) MTLC can support somedegree of (familiarity-based) recognition on its own.

This framework captures at a gross level how hippocampaldamage affects memory, but it is too vague to be useful inexplaining the considerable variability that exists across patientsand tests in how hippocampal damage affects recognition. Just assome hippocampal patients have shown more of a recognitiondeficit than others, some studies have found (within individualpatients) greater impairment on some recognition tests than others(see, e.g., Holdstock et al., 2002). In the absence of further spec-ification of the hippocampal contribution (and how it differs fromthe contribution of MTLC), it is not possible to proactively deter-mine whether recognition will be impaired or spared in a particularpatient and/or test.

Aggleton and Brown (1999) attempted to flesh out their theoryby arguing that MTLC familiarity can support recognition ofindividual items but that memory for new associations betweenitems depends on hippocampal recall (Eichenbaum, Otto, & Co-hen, 1994, and Sutherland & Rudy, 1989, made similar claims).This view implies that item recognition should be intact but theability to form new associations should be impaired after focalhippocampal damage. However, Andrew Mayes and colleagueshave found that hippocampally lesioned patient Y.R., who hasshown intact performance on some item-recognition tests (see,e.g., Mayes et al., 2002), showed impaired performance on otheritem-recognition tests (see, e.g., Holdstock et al., 2002) and sparedperformance on some tests that require participants to associatepreviously unrelated stimuli (e.g., the words window and reason;Mayes et al., 2001).

In summary, it is becoming increasingly evident that the effectsof hippocampal damage are complex. There appears to be somefunctional specialization in the medial temporal lobe, but thesimple dichotomies that have been proposed to explain this spe-cialization either are too vague (recall vs. familiarity) or areinconsistent with recently acquired data (item memory vs. memoryfor new associations).

Summary: Combining the Approaches

What should be clear at this point is that the math-modeling andcognitive-neuroscience approaches to recognition memory wouldgreatly benefit from increased cross talk: Math-modeling ap-proaches need a new source of constraints before they can fullyexplore how recall contributes to recognition, and cognitive-neuroscience approaches need a new, more mechanistically so-phisticated vocabulary for talking about the roles of different brainstructures to adequately characterize differences in how MTLCcontributes to recognition as compared with hippocampus.

The goal of our research is to achieve a synthesis of these twoapproaches by constructing a computational model of recognitionmemory in which there is a transparent mapping between differentparts of the model and different subregions of hippocampus andMTLC. This mapping makes it possible to address neuroscientificfindings using the model. For example, to predict the effects of aparticular kind of hippocampal lesion, we can lesion the corre-sponding region of the model. By bringing a wide range ofconstraints—both purely behavioral and neuroscientific—to bear

on a common set of mechanisms, we hope to achieve a moredetailed understanding of how recognition memory works.

Our model falls clearly in the dual-process tradition insofar aswe posit that the hippocampus and MTLC contribute signals withdistinct properties to recognition memory. The key, differentiatingproperty is that—in our model—differences in the two signals aregrounded in architectural differences between the hippocampusand MTLC; because most of these architectural differences fallalong a continuum, it follows that differences in the two signals aremore nuanced than the dichotomies (item vs. associative) dis-cussed above.

Precis of Modeling Work

In this section, we summarize the major claims of the article,with pointers to locations in the main text, below, where theseissues are discussed in greater detail.

Complementary Learning Systems

The hippocampus is specialized for rapidly memorizing specificevents, and the neocortex is specialized for slowly learning aboutthe statistical regularities of the environment. These are the centralclaims of the complementary-learning-systems (CLS) framework(McClelland et al., 1995; O’Reilly & Rudy, 2001). According tothis framework, the two goals of memorizing specific events andlearning about statistical regularities are in direct conflict whenimplemented in neural networks; thus, to avoid making a trade-off,human beings have evolved specialized neural systems for per-forming these tasks (see Marr, 1971; O’Keefe & Nadel, 1978;Sherry & Schacter, 1987, for similar ideas and Carpenter & Gross-berg, 1993, for a contrasting perspective).

The hippocampus assigns distinct ( pattern-separated) represen-tations to stimuli, thereby allowing it to learn rapidly withoutsuffering catastrophic interference. In contrast, neocortex assignssimilar representations to similar stimuli; use of overlapping rep-resentations allows neocortex to represent the shared structure ofevents and therefore makes it possible for neocortex to generalizeto novel stimuli as a function of their similarity to previouslyencountered stimuli.

A Dual-Process Model of Recognition

We have developed models of the hippocampus and neocortexthat incorporate key aspects of the biology of these structures andinstantiate the CLS principles outlined above (McClelland et al.,1995; O’Reilly & McClelland, 1994; O’Reilly & Munakata, 2000;O’Reilly, Norman, & McClelland, 1998; O’Reilly & Rudy 2001).(see Modeling Framework, below).

The cortical model supports familiarity judgments based on thesharpness of representations in MTLC. Competitive self-organization (arising from Hebbian learning and inhibitory com-petition) causes stimulus representations to become sharper overrepeated exposures (i.e., activity is concentrated in smaller numberof units; see The Cortical Model, below). However, the corticalmodel cannot support recall of details from specific events owingto its relatively low learning rate and its inability to sufficientlydifferentiate the representations of different events. Familiarity ismeasured by the activity of the top k most active units, though

613MODELING HIPPOCAMPAL AND NEOCORTICAL CONTRIBUTIONS

other measures are possible, and we have explored some of them,with similar results (for discussion of alternate measures, seeFuture Directions, below).

The hippocampal model supports recall of previously encoun-tered stimuli. When stimuli are presented at study, the hippocam-pal model assigns relatively nonoverlapping (pattern-separated)representations to these items in region CA3. Active units in CA3are linked to one another and to a copy of the input pattern (viaregion CA1). At test, presentation of a partial version of a studiedpattern leads to reconstruction ( pattern completion) of the originalCA3 representation and, through this, to reconstruction of theentire studied pattern on the output layer (see The HippocampalModel, below). The hippocampal model is applied to recognitionby computing the degree of match between retrieved informationand the recall cue, minus the amount of mismatch; recall ofmatching information is evidence that the cue was studied, andrecall of information that mismatches the retrieval cue is evidencethat the cue was not studied.

Part 1: Basic Network Properties

The signals generated by the hippocampal and cortical modelshave distinct operating characteristics. This difference is largelydue to differences in the two networks’ ability to carry out patternseparation (see Simulation 1: Pattern Separation and Simulation 2:Nature of the Underlying Distributions, below).

MTLC familiarity functions as a standard signal-detection pro-cess. In the cortical model, the familiarity distributions for studieditems and lures are Gaussian and overlap extensively. The distri-butions overlap because the representations of studied items andlures overlap in MTLC. Some lures, by chance, have representa-tions that overlap very strongly with the representations of studieditems in MTLC, and—as a result—these lures trigger a strongfamiliarity signal.

In contrast, the hippocampal recall signal is more diagnostic: Inthe hippocampal model, studied items sometimes trigger strongmatching recall, but most lures do not trigger any recall because ofthe hippocampus’s tendency to assign distinct representations tostimuli (regardless of similarity). As such, high levels of matchingrecall strongly indicate that an item was studied. However, thereare boundary conditions on the diagnosticity of the hippocampalrecall signal. When the average amount of overlap between stimuliis high, hippocampal pattern-separation mechanisms break down,resulting in strong recall of shared, prototypical features (even inresponse to lures) and in poor recall of features that are unique toparticular studied items.

The difference in operating characteristics between the MTLCfamiliarity and hippocampal recall signals is most evident onyes–no (YN) related-lure recognition tests where lures are similarto studied items but studied items are dissimilar to one another.The hippocampal model strongly outperforms the cortical modelon these tests (see Simulation 3: YN Related-Lure Simulations,below). As target–lure similarity increases, lure familiarity in-creases steadily, but (up to a point) hippocampal pattern separationworks to keep lure recall at floor. Very similar lures trigger recall,but when this happens the lure can often be rejected due tomismatch between retrieved information and the recall cue. Forexample, if the model studies rats and is tested with rat, it mightrecall having studied rats (and reject rat on this basis). On related-

lure recognition tests, the presence of any mismatching recall ishighly diagnostic of the item being a lure. As such, we apply arecall-to-reject strategy to hippocampal recall on such testswhereby items are given a confident new response if they triggerany mismatching recall.

Finally, in the Sources of Variability section, below, we discussdifferent kinds of variability (e.g., variability due to random initialweight settings, encoding variability) and their implications forrecognition performance.

Part 2: Applications to Behavioral Phenomena

Interactions Between Lure Relatedness and Test Format

The hippocampal model’s advantage for related lures (observedwith YN tests, as discussed above) is mitigated by giving themodels a forced choice between studied items and correspondingrelated lures at test (i.e., test A vs. A�, B vs. B�, where A� and B�are lures related to A and B, respectively; see Simulation 4:Lure-Relatedness and Test-Format Interactions, below). Corticalperformance benefits from the high degree of covariance in thefamiliarity scores triggered by studied items and correspondingrelated lures, which makes small familiarity differences highlyreliable. In contrast, use of this test format may actually harmhippocampal performance relative to other test formats. On forced-choice (FC) tests with noncorresponding lures (test A vs. B�, B vs.A�), the hippocampal model has two independent chances torespond correctly: On trials where it fails to recall the studied item(A), it can still respond correctly if the lure (B�) triggers mismatch-ing recall (and is rejected on this basis). Use of corresponding luresdeprives the model of this second chance insofar as recall triggeredby studied items and corresponding lures is highly correlated—ifA does not trigger recall, A� will not trigger recall-to-reject.

The predicted interaction between target–lure similarity and testformat was obtained in experiments comparing a focal hippocam-pal amnesic with control participants. The model predicts thathippocampally lesioned patients, who are relying exclusively onMTLC familiarity, should perform poorly relative to controls onstandard YN recognition tests with related lures, but they shouldperform relatively well on FC tests with corresponding relatedlures, and they should perform well relative to controls on both YNand FC tests that use unrelated lures (insofar as both networksdiscriminate well between studied items and unrelated lures).Holdstock et al. (2002) and Mayes et al. (2002) found exactly thispattern of results in hippocampally lesioned patient Y.R.

Associative Recognition

Associative recognition tests (i.e., study A–B, C–D; test withstudied pairs and recombined pairs like A–D) can be viewed as aspecial case of the related-lure paradigm discussed above (seeSimulation 5: Associative Recognition and Sensitivity to Conjunc-tions, below). The hippocampal model outperforms the corticalmodel on YN associative-recognition tests. As in the related-luresimulations, the hippocampal advantage is due to hippocampalpattern separation, and the hippocampus’s ability to carry outrecall-to-reject. However, even though the cortical model is worseat associative recognition than the hippocampus, the cortical modelstill performs well above chance. This finding shows that our


cortical model has some (albeit reduced) sensitivity to whetheritems occurred together at study, unlike other models (e.g., Rudy& Sutherland, 1989) that posit that cortex supports memory forindividual features but not novel feature conjunctions. We alsofound that giving the models a forced choice between studied pairsand overlapping re-paired lures (test A–B vs. A–D) mitigates thehippocampal advantage for associative recognition. This mirrorsthe finding that FC testing with corresponding related lures miti-gates the hippocampal advantage for related lures. Finally, wepresent data from previously published studies that support themodel’s predictions regarding how focal hippocampal damageshould affect associative-recognition performance, as a function oftest format.

Interference Effects

Hippocampally and cortically driven recognition can be differ-entially affected by interference manipulations: Increasing liststrength impairs discrimination of studied items and lures in thehippocampal model but does not impair discrimination based onMTLC familiarity. The list-strength paradigm measures how re-peated study of a set of interference items affects participants’ability to discriminate between nonstrengthened (but studied) tar-get items and lures (see Simulation 6: Interference and ListStrength, below).

In both models, the overall effect of interference is to decreaseweights to discriminative features of studied items and lures and toincrease weights to prototypical features (which are shared by ahigh proportion of items in the item set). The hippocampal modelpredicts a list-strength effect (LSE) because increasing list strengthreduces recall of discriminative features of studied items, and lurerecall is already at floor. Increasing list strength therefore has theeffect of pushing the studied and lure recall distributions together(reducing discriminability). The cortical model predicts a null LSEbecause the familiarity signal triggered by lures has room todecrease as a function of interference. Increasing list strengthreduces responding to (the discriminative features of) both studieditems and lures, but the average difference in studied and lurefamiliarity does not decrease, so discriminability does not suffer.

In the cortical model, lure familiarity initially decreases morethan studied-item familiarity as a function of list strength, so thestudied–lure gap in familiarity actually widens slightly with in-creasing interference. The widening of the studied–lure gap can beexplained in terms of differentiation: Studying an item makes itsrepresentation overlap less with the representations of other, inter-fering items (Shiffrin, Ratcliff, & Clark, 1990); therefore, studieditems suffer less interference than lures. However, according to themodel, there are limits on this dynamic. With high levels ofinterference item strengthening and/or high input overlap, thecortical model’s sensitivity to discriminative features of studieditems and lures approaches floor, and the studied and lure famil-iarity distributions start to converge (resulting in decreaseddiscriminability).

Data from two recent list-strength experiments provide supportfor the model’s prediction that list strength should affect recall-based discrimination but not familiarity-based discrimination(Norman, 2002). We also discuss ways of modifying the learningrule to accommodate the finding that increasing list length (i.e.,adding new items to the study list) hurts recognition sensitivity

more than increasing list strength (Murnane & Shiffrin, 1991a; seeList-Length Effects, below).

The Combined Model and Independence

The extent to which the neocortical and hippocampal recogni-tion signals are correlated varies in different conditions. Theseissues are explored using a more realistic combined model wherethe cortical network responsible for computing familiarity servesas the input to the hippocampal network (see Simulation 7: TheCombined Model and the Independence Assumption, below).

In the combined model, variability in how well items are learnedat study (encoding variability) bolsters the correlation betweenrecall and familiarity signals, as was postulated by Curran andHintzman (1995) and others. In contrast, increasing interferencereduces the recall–familiarity correlation for studied items. Thisoccurs because interference pushes raw recall and familiarityscores in different directions (increasing interference reduces re-call, but asymptotically it boosts familiarity by boosting the mod-el’s responding to shared, prototypical-item features). Taken to-gether, these results show that recall and familiarity can beindependent when there is enough interference to counteract theeffects of encoding variability.

Effects of Partial Lesions

Partial hippocampal lesions can lead to worse overall recogni-tion performance than complete lesions (see Simulation 8: LesionEffects in the Combined Model, below). In the hippocampal model,partial lesions cause pattern-separation failure, which sharply re-duces the diagnosticity of the recall signal. Recognition perfor-mance suffers because the noisy recall signal drowns out usefulinformation that is present in the familiarity signal. Moving froma partial hippocampal lesion to a complete lesion improves per-formance by removing this source of noise. In contrast, increasingMTLC lesion size in the model leads to a monotonic decrease inperformance. This occurs because MTLC lesions directly impairfamiliarity-based discrimination and indirectly impair recall-baseddiscrimination (because MTLC serves as the input to the hip-pocampus). These results are consistent with a recent meta-analysis of the lesion data showing a negative correlation betweenrecognition impairment and hippocampal lesion size and a positivecorrelation between recognition impairment and MTLC lesion size(Baxter & Murray, 2001b).

A Note on Terminology

We use the terms recall and familiarity to describe the respec-tive contributions of the hippocampus and MTLC to recognitionmemory because these terms are heuristically useful. The hip-pocampal contribution to recognition is recall insofar as it involvesretrieval of specific studied details. We use familiarity to describethe MTLC signal because it adheres to the definition of familiarityset forth by Hintzman (1988), Gillund and Shiffrin (1984), andothers, that is, it is a scalar that tracks the global match orsimilarity of the test probe to studied items.

However, we realize that the terms recall and familiarity comewith a substantial amount of theoretical baggage. Over the years,researchers have made a very large number of claims regarding


properties of recall and familiarity; for example, Yonelinas hasargued that recall is a high-threshold process (see, e.g., Yonelinas,2001), and Mandler (1980) and others (e.g., Aggleton & Brown,1999) have argued that familiarity reflects memory for individualitems apart from their associations with other items and contexts.By linking the hippocampal contribution with recall and theMTLC contribution with familiarity, we do not mean to say that allof the various (and sometimes contradictory) properties that havebeen ascribed to recall and familiarity over the years apply to thehippocampal and MTLC contributions, respectively. Just the op-posite: We hope to redefine the properties of recall and familiarityusing neurobiological data on the properties of the hippocampusand MTLC. In this article, we systematically delineate how theCLS model’s claims about hippocampal recall and MTLC famil-iarity deviate from claims made by existing dual-process theories.

Modeling Framework

Both the hippocampal and neocortical networks utilize the Heb-bian component of O’Reilly’s Leabra algorithm (O’Reilly, 1996,1998; O’Reilly & Munakata, 2000; the full version of Leabra alsoincorporates error-driven learning, but error-driven learning wasturned off in the simulations reported here). The algorithm we usedincorporates several widely accepted characteristics of neural com-putation, including Hebbian long-term potentiation/long-term de-pression (LTP/LTD) and inhibitory competition between neurons,that were first brought together by Grossberg (1976). (For moreinformation on these mechanisms, see also Kanerva, 1988; Ko-honen, 1977; Minai & Levy, 1994; Oja, 1982; Rumelhart &Zipser, 1986.) In our model, LTP is implemented by strengtheningthe connection (weight) between two units when both the sendingand receiving units are active together; LTD is implemented byweakening the connection between two units when the receivingunit is active but the sending unit is not (heterosynaptic LTD).Inhibitory competition is implemented using a k-winners-take-all(kWTA) algorithm, which sets the amount of inhibition for agiven layer such that at most k units are strongly active. Al-though the kWTA rule sets a firm limit on the number of unitsthat show strong (�.25) activity, there is still considerable flexi-bility in the overall distribution of activity across units in a layer.This is important for our discussion of sharpening in the corticalmodel, below. Key aspects of the algorithm are summarized inAppendix A.

The Cortical Model

The cortical network is composed of two layers, input (whichcorresponds to cortical areas that feed into MTLC) and hidden(corresponding to MTLC; see Figure 2). The basic function of themodel is for the hidden layer to encode regularities that are presentin the input layer; this is achieved through the Hebbian learningrule. To capture the idea that the input layer represents manydifferent cortical areas, it consists of twenty-four 10-unit slots,with 1 unit out of 10 active in each slot. A useful way to think ofslots is that different slots correspond to different feature dimen-sions (e.g., color or shape) and different units within a slot corre-spond to different, mutually exclusive features along that dimen-sion (e.g., shapes: circle, square, triangle). The hidden (MTLC)layer consists of 1,920 units, with 10% activity (i.e., 192 of these

units are active on average for a given input). The input layer isconnected to the MTLC layer via a partial feedforward projectionwhere each MTLC unit receives connections from 25% of theinput units. When items are presented at study, these connectionsare modified via Hebbian learning.

Input patterns were constructed from prototypes by randomlyselecting a feature value (possibly identical to the prototype featurevalue) for a random subset of slots. The number of slots that areflipped (i.e., given a random value) when generating items fromthe prototype is a model parameter—increasing the number ofslots that are flipped decreases the average overlap between items.When all 24 slots are flipped, the resulting item patterns have 10%overlap with one another on average (i.e., exactly as expected bychance in a layer with a 10% activation level). Thus, with inputpatterns, one can make a distinction between prototypical featuresof those patterns, which have a relatively high likelihood of beingshared across input patterns, and nonprototypical, item-specificfeatures of those patterns (generated by randomly flipping slots),which are relatively less likely to be shared across input patterns.Prototype features can be thought of as representing both high-frequency item features (e.g., if one studies pictures of people fromNorway, one sees that most people there have blond hair) as wellas contextual features that are shared across multiple items in anexperiment (e.g., the fact that all of the pictures are viewed on aparticular monitor in a particular room on a particular day). Somesimulations involved more complex stimulus construction, as de-scribed where applicable.

To apply the cortical model to recognition, we exploited the factthat—as items are presented repeatedly—their representations inthe MTLC layer become sharper (see Figure 3). That is, novelstimuli weakly activate a large number of MTLC units, whereasfamiliar (previously presented) stimuli strongly activate a rela-tively small number of units. Sharpening occurs because Hebbianlearning specifically tunes some MTLC units to represent thestimulus. When a stimulus is first presented, some MTLC units by

Figure 2. Diagram of the cortical network. The cortical network consistsof two layers, an input layer (corresponding to lower cortical regions thatfeed into medial temporal lobe cortex [MTLC]) and a hidden layer (cor-responding to MTLC). Units in the hidden layer compete to encode (viaHebbian learning) regularities that are present in the input layer.


chance respond more strongly to the stimulus than other units;these units get tuned by Hebbian learning to respond even morestrongly to the item the next time it is presented, and these stronglyactive units start to inhibit units that are less strongly active (foradditional discussion of the idea that familiarization causes someunits to drop out of the stimulus representation, see Li, Miller, &Desimone, 1993). We should note that sharpening is not a novelproperty of our model—rather, it is a general property ofcompetitive-learning networks with graded unit-activation valuesin which there is some kind of contrast enhancement within a layer(see, e.g., Grossberg, 1986, Section 23; Grossberg & Stone, 1986,Section 16).

The sharpening dynamic in our model is consistent with neuraldata on the effects of repeated presentation of stimuli in cortex.Electrophysiological studies have shown that some neurons thatinitially respond to a stimulus exhibit a lasting decrease in firingwhereas other neurons that initially respond to the stimulus do notexhibit decreased firing (see, e.g., Brown & Xiang, 1998; Li et al.,1993; Miller, Li, & Desimone, 1991; Riches, Wilson, & Brown,1991; Rolls, Baylis, Hasselmo, & Nalwa, 1989; Xiang & Brown,1998). According to our model, the latter population consists ofneurons that were selected (by Hebbian learning) to represent thestimulus, and the former population consists of neurons that arebeing forced out of the representation via inhibitory competition.

To index representational sharpness in our model—and throughthis, stimulus familiarity—we measured the average activity of theMTLC units that won the competition to represent the stimulus.That is, we took the average activation of the top k (192 or 10% ofthe MTLC) units computed according to the kWTA inhibitory-competition function. This activation of winners (act win) mea-sure increases monotonically as a function of how many times astimulus was presented at study. In contrast, the simpler alternativemeasure of using the average activity of all units in the layer is notguaranteed to increase as a function of stimulus repetition—as astimulus becomes more familiar, the winning units become moreactive, but losing units become less active (due to inhibition from

the winning units); the net effect is therefore a function of the exactbalance between these increases and decreases (for an example ofanother model that bases recognition decisions on an activityreadout from a neural network doing competitive learning, seeGrossberg & Stone, 1986).

Although we used act win in the simulations reported below,we do not want to make a strong claim that act win is the way thatfamiliarity is read out from MTLC. It is the most convenient andanalytically tractable way to do this in our model, but it is far fromthe only way of operationalizing familiarity, and it is unclear howother brain structures might read out act win from MTLC. Webriefly describe another, more neurally plausible familiarity mea-sure (the time it takes for activation to spread through the network)in the General Discussion section.

Finally, we should point out that the idea (espoused above) thatthe same network is involved in feature extraction and familiaritydiscrimination is controversial; in particular, Malcolm Brown,Rafal Bogacz, and their colleagues (see, e.g., Bogacz & Brown,2003; Brown & Xiang, 1998) have argued that specialized popu-lations of neurons in MTLC are involved in feature extractionversus familiarity discrimination. At this point, it suffices to saythat our focus in this article is on the familiarity-discriminationcapabilities of the cortical network rather than its ability to extractfeatures. We address Brown and Bogacz’s claims in more detail inthe General Discussion.

The Hippocampal Model

We have developed a standard model of the hippocampus(O’Reilly & Munakata, 2000; O’Reilly et al., 1998; O’Reilly &Rudy, 2001; Rudy & O’Reilly, 2001) that implements widelyaccepted ideas of hippocampal function (Hasselmo, 1995; Hebb,1949; Marr, 1971; McClelland et al., 1995; McNaughton & Mor-ris, 1987; O’Reilly & McClelland, 1994; Rolls, 1989). Our goal inthis section is to describe the model in just enough detail tomotivate the model’s predictions about recognition memory. Ad-ditional details regarding the architecture of the hippocampalmodel (e.g., the percentage activity in each layer of the model) areprovided in Appendix B.

In the brain, entorhinal cortex (EC) is the interface betweenhippocampus and neocortex; superficial layers of EC send input tothe hippocampus, and deep layers of EC receive output from thehippocampus (see Figure 1). Correspondingly, our model subdi-vides EC into an EC in layer that sends input to the hippocampusand an EC out layer that receives output from the hippocampus.Like the input layer of the cortical model, both EC in and EC outhave a slotted structure (twenty-four 10-unit slots, with 1 unit perslot active).

Figure 4 shows the structure of the model. The job of thehippocampal model, stated succinctly, is to store patterns of EC inactivity in a manner that supports subsequent recall of thesepatterns on EC out. The hippocampal model achieves this goal inthe following stages: Input patterns are presented to the model byclamping those patterns onto the input layer, which serves toimpose the pattern on EC in via fixed, one-to-one connections.From EC in, activation spreads both directly and indirectly (viathe dentate gyrus [DG]) to region CA3. The resulting pattern ofactivity in CA3 is stored by Hebbian weight changes in thefeedforward pathway and by strengthening recurrent connections

Figure 3. Illustration of the sharpening of hidden (medial temporal lobecortex [MTLC]) layer activation patterns in a miniature version of ourcortical model. A: The network prior to sharpening; MTLC activations(more active � lighter color) are relatively undifferentiated. B: The net-work after Hebbian learning and inhibitory competition produce sharpen-ing; a subset of the units are strongly active, with the remainder inhibited.In this example, we would read out familiarity by measuring the averageactivity of the k � 5 most active units.


in CA3 between active units; these weight changes serve to bindthe disparate elements of the input pattern by linking them to ashared episodic representation in CA3.

An important property of DG and CA3 is that representations inthese structures are very sparse—relatively few units are active fora given stimulus. The hippocampus’s use of sparse representationsgives rise to pattern separation. If only a few units are active perinput pattern, then overlap between the hippocampal representa-tions of different items tends to be minimal (Marr, 1971; seeO’Reilly & McClelland, 1994, for a mathematical analysis of howsparseness results in pattern separation and the role of the DG infacilitating pattern separation).

Next, to complete the loop, the CA3 representation needs to belinked back to the original input pattern. This is accomplished bylinking the CA3 representation to active units in region CA1. LikeCA3, region CA1 contains a re-representation of the input pattern.However, unlike the CA3 representation, the CA1 representation isinvertible—if an item’s representation is activated in CA1, well-established connections between CA1 and EC out allow activityto spread back to the item’s representation in EC out. Thus, CA1serves to translate between sparse representations in CA3 andmore overlapping representations in EC (for more discussion ofthis issue, see McClelland & Goddard, 1996; O’Reilly et al.,1998).

At test, when a previously studied EC in pattern (or a subsetthereof) is presented to the hippocampal model, the model is

capable of reactivating the entire CA3 pattern corresponding tothat item via strengthened weights in the EC-to-CA3 pathway andstrengthened CA3 recurrents. Activation then spreads from theitem’s CA3 representation to the item’s CA1 representation viastrengthened weights and (from there) to the item’s EC out rep-resentation. In this manner, the hippocampus manages to retrievea complete version of the studied EC pattern in response to apartial cue.

To apply the hippocampal model to recognition, we exploitedthe fact that studied items tend to trigger recall (of the item itself),more so than lure items. Thus, a high level of match between thetest probe (presented on the EC input layer) and recalled informa-tion (activated over the EC output layer) constitutes evidence thatan item was studied. Also, we exploited the fact that lures some-times trigger recall of information that mismatches the recall cue.Thus, mismatch between recalled information and the test probetends to indicate that an item was not studied.

For each test item, we generated a recall score using the formula

�match � mismatch�/�numslots�, (1)

where match is the number of recalled features (on EC out) thatmatch the cue (on EC in), and mismatch is likewise the numberthat mismatch (a feature is counted as recalled if the unit corre-sponding to that feature in EC out shows activity � .9); numslotsis a constant that reflects the total number of feature slots in EC(24, in these simulations).

One should appreciate that Equation 1 is not the only way toapply the hippocampal model to recognition. For example, insteadof looking at recall of the test cue itself, one could attach contex-tual tags to items at study, leave these tags out at test, and measurethe extent to which items elicit recall of contextual tags. Also, thisequation does not incorporate the fact that recall of item-specificfeatures (i.e., features unique to particular items in the item set) ismore diagnostic of study status than recall of prototypical fea-tures—if all items in the experiment are fish, recall of prototypicalfish features (e.g., I studied fish) in conjunction with a test itemdoes not mean that one studied this particular item. We selected thematch � mismatch rule because it is a simple way to reduce thevector output of the hippocampal model to a scalar that correlateswith the study status of test items. Assessing the optimality of thisrule relative to other rules and exploring ways in which differentrules might be implemented neurally are topics for future research.

The only place where we deviate from using match � mismatchin this article is in our related-lure simulations, where distractorsare related to particular studied items but studied items are rea-sonably distinct from one another. In this situation, we use therecall-to-reject rule that places a stronger weight on mismatchingrecall. The Simulation 3: YN Related-Lure Simulations section ofthe article, below, contains a detailed account of our reasons forusing recall-to-reject and the implications of using this rule inplace of match � mismatch.

Simulation Methodology

Our initial simulations involved a side-by-side comparison of the corti-cal and hippocampal networks receiving the exact same input patterns. Thismethod allowed us to analytically characterize differences in how thesenetworks responded to stimuli. A shortcoming of this side-by-side ap-proach is that we could not explore direct interactions between the two

Figure 4. Diagram of the hippocampal network. The hippocampal net-work links input patterns in entorhinal cortex (EC) to relatively nonover-lapping (pattern-separated) sets of units in region CA3; recurrent connec-tions in CA3 bind together all of the units involved in representing aparticular EC pattern; the CA3 representation is linked back to EC viaregion CA1. Learning in the CA3 recurrent connections and in projectionslinking EC to CA3 and CA3 to CA1 makes it possible to recall entire storedEC patterns on the basis of partial cues. The dentate gyrus (DG) serves tofacilitate pattern separation in region CA3.


systems. To remedy this shortcoming, we also present simulations using acombined model where the cortical and hippocampal networks are con-nected in serial (such that the cortical regions involved in computingstimulus familiarity serve as the input to the hippocampal network)—thisarrangement more accurately reflects how cortex and hippocampus arearranged in the brain (see Figure 1).

Basic Method

In our recognition simulations, for each simulated participant we reran-domized the connectivity patterns for partial projections (e.g., if the spec-ified amount of connectivity between Layer X and Layer Y was 25%, theneach unit in Layer Y was linked at random to 25% of the units in Layer X),and we initialized the network weights to random values (weight values inthe cortical model were sampled from a uniform distribution with mean �.5 and range � .25; see Appendix B for weight initialization parameters fordifferent parts of the hippocampal model). The purpose of this randomiza-tion was to ensure that units in MTLC and the hippocampus would developspecialized receptive fields (i.e., they would be more strongly activated bysome input features than others)—this kind of specialization is necessaryfor competitive learning.

After randomization was complete, the cortical and hippocampal modelswere (separately) given a list of items to learn, followed by a recognitiontest in which the models had to discriminate between 10 studied targetitems and 10 nonstudied lure items. No learning occurred at test. Unlessotherwise specified, all of our recognition simulations used the same set ofparameters (hereafter referred to as the basic parameters; these parametersare described in detail in Appendix C). In our basic-parameter simulations,we used a 20-item study list (10 target items plus 10 interference items thatwere not tested), and the average amount of overlap between items was20%—20% overlap between items was achieved by starting with a 24-slotprototype pattern and then generating items by randomly selecting a featurevalue for 16 randomly selected slots (note that each item overlapped atleast 33% with the prototype but, on average, items overlapped 20% witheach other).

To facilitate comparison between the models, we used hippocampal andcortical parameters that yielded roughly matched performance across thetwo models for both single-probe (YN) and FC recognition tests. Wematched performance in this way to alleviate concerns that differentialeffects of manipulations on hippocampal recall and MTLC familiaritymight be attributable simply to different overall levels of performance inthe two networks. However, this matching does not constitute a strongclaim that hippocampal and cortical performance are—in reality—matchedwhen overlap equals 20% and study-list length equals 20.

Simulating YN and FC Testing

To simulate YN recognition performance, items were presented one at atime at test, and we recorded the familiarity score (in the cortical model)and recall score (in the hippocampal model) triggered by each item. For thecortical model, we set an unbiased criterion for each simulated participantby computing the average familiarity scores associated with studied andlure items, respectively, and then placing the familiarity criterion exactlybetween the studied and lure means. All items triggering familiarity scoresabove this criterion were called old.

For the hippocampal model, we took a different approach to criterionsetting; as discussed in Simulation 2: Nature of the Underlying Distribu-tions, below, it is possible to set a high recall criterion that is sometimescrossed by studied items but never crossed by lures. We assumed thatparticipants would be aware of this fact (i.e., that high amounts of recall areespecially diagnostic of having studied an item) and set a recall criterionthat was high enough to avoid false recognition. Accordingly, in ourbasic-parameter simulations, we used a fixed, relatively high criterion forcalling items old (recall � .40). This value was chosen because—assuming

other parameters were set to their basic values—it was sometimes ex-ceeded by studied items but never by lures (unless lures were constructedto be similar to specific studied items; see Simulation 3: YN Related-LureSimulations, below, for more details).

For both models, we used d� (computed on individual participants’ hitand false-alarm rates) to index YN recognition sensitivity.1

Our method for simulating FC testing was straightforward: We pre-sented the two test alternatives one at a time. For the cortical model, werecorded the act win score associated with each test alternative and se-lected the item with the higher score. For the hippocampal model, werecorded the match � mismatch score associated with each test alternativeand selected the item with the higher score. For both models, if there wasa tie, one of the two test alternatives was selected at random.

All of the simulation results reported in the text of this article aresignificant at p � .001. In graphs of simulation results (starting withSimulation 3: YN Related-Lure Simulations, below), error bars indicate thestandard error of the mean computed across simulated participants. Whenerror bars are not visible, this is because they are too small relative to thesize of the symbols on the graph (and thus are covered by the symbols).

Part 1: Basic Network Properties

Simulations reported in this section addressed basic propertiesof the cortical and hippocampal networks, including differences intheir ability to assign distinct (pattern-separated) representations toinput patterns and differences in their operating characteristics. Wealso discuss sources of variability in the two networks.

Simulation 1: Pattern Separation

Many of the differences between hippocampally and corticallydriven recognition in our model arise from the fact that the hip-pocampal network exhibits more pattern separation than the cor-tical network. To document the two networks’ pattern-separationabilities, we ran simulations where we manipulated the amount ofoverlap between paired input patterns. The first item in each pairwas presented at study, and the second item was presented at test.For each pair, we measured the resulting amount of overlap inregion CA3 of the hippocampal model and in the hidden (MTLC)layer of the cortical model. Pattern separation is evident whenoverlap between the internal representations of paired items (inCA3 or MTLC) is less than the amount of input overlap.

The results of these simulations (see Figure 5) confirm that,although both networks show some pattern separation, the amountof pattern separation is larger in the hippocampal model. They alsoshow that the hippocampus’s ability to assign distinct representa-tions to stimuli is limited—as overlap between input patternsincreases, hippocampal overlap eventually increases above floorlevels (although it always lags behind input-pattern overlap).

Simulation 2: Nature of the Underlying Distributions

One way to characterize how cortical and hippocampal contri-butions to recognition differ is to plot the distributions of these

1 d� � z(H) � z(F), where z is the inverse of the normal distributionfunction, H is the hit rate, and F is the false-alarm rate. To avoid problemswith d� being undefined when hit or false-alarm rates equalled 0 or 1, weadjusted hit and false-alarm rates using the correction suggested bySnodgrass and Corwin (1988) prior to computing d�: P � (n � 5)/(N � 1),where n is the number of old responses, N is the total number of items, andP is the corrected-percentage old value.


signals for studied items and lures. Given 20% average overlapbetween input patterns, the MTLC familiarity distributions forstudied items and lures are Gaussian and overlap strongly (seeFigure 6A)—this is consistent with the idea, expressed by Yoneli-nas, Dobbins, Szymanski, Dhaliwal, and King (1996) and manyothers (e.g., Hintzman, 1988), that familiarity is well described bystandard (Gaussian) signal-detection theory. In contrast, the hip-pocampal recall distributions (see Figure 6B) do not adhere to asimple Gaussian model. The bulk of the lure recall distribution islocated at the zero recall point, although some lures trigger above-

zero recall. The studied recall distribution is bimodal, and cru-cially, it extends further to the right than the lure recall distribu-tion, so there are some (high) recall scores that are sometimestriggered by studied items but never by lures. Thus, high levels ofrecall are highly diagnostic—for these parameters, if an itemtriggers a recall score of .2 or greater, one can be completely surethat it was studied.

The low overall level of lure recall in this simulation can beattributed to hippocampal pattern separation. Because of patternseparation, the CA3 representations of lures do not overlapstrongly with the CA3 representations of studied items. Becausethe CA3 units activated by lures (for the most part) were notactivated at study, these units do not possess strong links to CA1;as such, activity does not spread from CA3 to CA1, and recall doesnot occur.

The studied recall distribution is bimodal because of nonlinearattractor dynamics in the hippocampus. If a studied test cue ac-cesses a sufficiently large number of strengthened weights, ittriggers pattern completion: Positive feedback effects (e.g., in theCA3 recurrents) result in strong reactivation of the CA3 and CA1units that were activated at study, thereby boosting recall. Moststudied items benefit from these positive feedback effects, but,because of variability in initial weight values, some studied itemsdo not have weights strong enough to yield positive feedback.These items only weakly activate CA3 and are poorly recalled,thereby accounting for the extra peak at recall � 0.

Increasing the average amount of overlap between items reducesthe diagnosticity of the hippocampal recall signal. When theamount of overlap between input patterns is high (e.g., 40.5%instead of 20%), both studied items and lures trigger large amountsof recall, such that the studied and lure recall distributions areroughly Gaussian and overlap extensively (see Figure 7).

High levels of lure recall occur in the high-overlap conditionbecause of pattern-separation failure in the hippocampus. As doc-umented in Simulation 1 (Figure 5), the hippocampus loses itsability to assign distinct representations to input patterns whenoverlap between inputs is very high. In this situation, the sameCA3 units—the units that are most sensitive to frequently occur-ring prototype features—are activated again and again by studied

Figure 5. Results of simulations exploring pattern separation in thehippocampal and cortical models. In these simulations, we created pairs ofitems and manipulated the amount of overlap between paired items. Thegraph plots the amount of input-layer overlap for paired items versus (a)CA3 overlap in the hippocampal model and (b) medial temporal lobecortex (MTLC) overlap in the cortical model. All points below the diagonal(dashed line) indicate pattern separation (i.e., representational overlap �input overlap). The hippocampal model shows a strong tendency towardpattern separation (CA3 overlap � � input overlap); the cortical modelshows a smaller tendency toward pattern separation (MTLC overlap isslightly less than input overlap). Hippo � hippocampus.

Figure 6. A: Histogram of the studied and lure medial temporal lobe cortex (MTLC) familiarity distributionsfor 20% average overlap. B: Histogram of the studied and lure hippocampal recall distributions for 20% averageoverlap.


patterns, and these units acquire very strong weights to the repre-sentations of prototype features in CA1. When items are presentedat test, they activate these core CA3 units to some extent (regard-less of whether or not the test item was studied), and activationspreads very quickly to CA1, leading to possibly erroneous recallof prototype features in response to both studied items and lures.Figure 8 shows that increasing overlap increases the probabilitythat prototypical features of studied items and lures will be recalledand reduces the probability that item-specific features of studieditems will be recalled (in part, because the hippocampus intrudesprototypical features in place of these item-specific features).

In summary, the results presented here show that hippocampalrecall has two modes of operation: When input patterns have lowto moderate average overlap, high levels of matching recall arehighly diagnostic—studied items sometimes trigger strong recall(of item-specific and prototype features), but lures trigger virtuallyno recall. In contrast, when input patterns have high averageoverlap, recall functions as a standard signal-detection process—

both studied items and lures trigger varying degrees of prototyperecall.

Simulation 3: YN Related-Lure Simulations

The strengths of the hippocampal model are most evident on YNrelated-lure recognition tests, where lures are similar to studieditems but studied items are dissimilar to one another. In thissection, we show how the hippocampal model outperforms thecortical model on these tests because of its superior pattern-separation and pattern-completion abilities.

Method. To simulate the related-lure paradigm, we first createdstudied-item patterns with 20% average overlap between items. Then, foreach studied (target) item, we created a related-lure item by taking thestudied item and flipping a prespecified number of slots; to vary target–luresimilarity, we varied the number of slots that we flipped to generate lures(less flipping resulting in more overlap). For comparison, we also ransimulations with unrelated lures that were sampled from the same itempool as studied items.

In the related-lure simulations presented here (and later in the article),we used a recall-to-reject hippocampal decision rule instead of our standardmatch � mismatch rule. According to this rule, items that trigger anymismatch are given a new response, otherwise the decision is based onmatch. We used this rule for two reasons: First, there is extensive empiricalevidence that, when lures are similar to studied items but studied items areunrelated to one another, participants use recall-to-reject (see, e.g., Rotello& Heit, 2000; Rotello, Macmillan, & Van Tassel, 2000; Yonelinas, 1997;but see Rotello & Heit, 1999, for a contrasting view). Second, it iscomputationally sensible to use recall-to-reject—we show that the presenceof mismatching recall in this paradigm is highly diagnostic of an item beingnonstudied.

Results. Figure 9 shows the results of our related-lure simula-tions: Recognition performance based on MTLC familiarity getssteadily worse as lures become increasingly similar to studieditems; in contrast, recognition based on hippocampal recall isrelatively robust to the lure-similarity manipulation. Figure 9 alsoshows that hippocampal performance is somewhat better for re-

Figure 7. Histogram of the studied and lure hippocampal recall distribu-tions for 40.5% average overlap.

Figure 8. Plot of the probability that item-specific and prototypicalfeatures of studied items and lures will be recalled, as a function of overlap.As overlap increases, the amount of prototype recall triggered by studieditems and lures increases, and the amount of item-specific recall triggeredby studied items decreases.

Figure 9. Yes–no (YN) recognition sensitivity (d�) in the two models, asa function of target–lure similarity. Target–lure similarity was operation-alized as the proportion of input features shared by targets and correspond-ing lures; note that the average level of overlap between studied (target)items was held constant at 20%. These simulations show that the hip-pocampal model is more robust to increasing target–lure similarity than thecortical model. The figure also shows that hippocampal performance forrelated lures is better when a recall-to-reject rule is used instead of match� mismatch. Hippo � hippocampus; MTLC � medial temporal lobecortex.


lated lures when the recall-to-reject rule is used (as opposed tomatch � mismatch).

The cortical model results can be explained in terms of the factthat the cortical model assigns similar representations to similarstimuli—because the representations of similar lures (vs. dissim-ilar lures) overlap more with the representations of studied items,similar lures benefit more from learning that occurred at study.Thus, lure familiarity smoothly tracks target–lure similarity; in-creasing similarity monotonically lowers the target–lure familiar-ity difference, leading to decreased discriminability.

In contrast, the hippocampal recall signal triggered by lures isstuck at floor until target–lure similarity is greater than 60%, andlures do not start to trigger above-criterion (i.e., � .40) recall untiltarget–lure similarity is greater than 80%. This occurs because ofhippocampal pattern separation—lures have to be very similar tostudied items before they access enough strengthened weights totrigger recall.

The hippocampus also benefits from the fact that lures some-times trigger pattern completion of the corresponding studied itemand can subsequently be rejected based on mismatch betweenrecalled information and the test cue. Figure 10 illustrates the point(mentioned earlier) that when lures resemble studied items butstudied items are not related to one another, mismatching recall ishighly diagnostic—studied items virtually never trigger mismatch-ing recall, but lures sometimes do. As such, it makes sense to usea rule (like recall-to-reject) that assigns a very high weight tomismatching recall.

Figure 10 also shows that mismatching recall triggered by luresincreases substantially with increasing target–lure similarity. Thisincrease in mismatching recall helps offset, to some degree, in-creased matching recall triggered by related lures. With recall-to-reject, the only way that lures can trigger an old response is if theytrigger a large amount of matching recall but no mismatchingrecall. The odds of this happening are very low.

In summary, these simulations demonstrate that both networkscan support good performance on YN recognition tests with luresthat are unrelated to studied items. When the networks are chal-lenged by boosting target–lure similarity, performance in bothnetworks suffers; however, the hippocampus is more robust to thismanipulation than cortex. As such, the model predicts that recog-

nition discrimination based on hippocampal recall should be betterthan discrimination based on MTLC familiarity on YN tests withrelated lures. This prediction is consistent with the view, expressedin several empirical studies, that recall is especially important fordiscriminating between studied items and very similar distractors(see, e.g., Hintzman, Curran, & Oppy, 1992; Rotello et al., 2000).

Sources of Variability

The final issue that we need to address in this Part 1: BasicNetwork Properties section is variability. Recognition performanceinvolves detecting the presence of a variable memory signalagainst a background of noise. Many distinct forms of variabilitycan affect recognition performance; we need to carefully delineatewhich of these sources of variability are present in our modelsbecause—as we show later—different forms of variability havedifferent implications for recognition performance.

The primary source of variability in our models is samplingvariability: variation in how well, on average, neural units areconnected to (sampled by) other neural units in the network. Notethat our use of the term sampling variability differs from how othermodelers have used this term. In our model, sampling variability isa function of variability in the initial values assigned to weights inthe network. Other models use sampling variability to refer tovariability in which item features are presented to the model atstudy and test (Atkinson & Estes, 1963) or variability in whichmemory trace is retrieved at test (Gillund & Shiffrin, 1984).

Sampling variability arises because, at the beginning of eachsimulation, weight strengths and connectivity patterns are set ran-domly. As discussed earlier, this randomness helps units in MTLCand the hippocampus form specialized representations of the inputspace. It also has the consequence that, by chance, some inputfeatures are sampled better (by units further downstream) thanother input features.

An important property of sampling variability is that it decreasesas network size increases. Intuitively, as the number of units andconnections increases, the odds that any one input pattern will beundersampled relative to another decreases. We conducted simu-lations to explore this issue, and the results are very clear: As weincrease network size, variability in MTLC-familiarity andhippocampal-recall scores steadily decreases, and d� scoressteadily increase.

In a network scaled to the approximate size of the human brain,sampling variability would likely be negligible. Therefore, weconclude that other forms of variability must be at play in thehuman brain; we briefly describe some other sources of variabilitybelow.

Other Sources of Variability

A potentially important source of variability in recall and fa-miliarity scores is variability in how well stimuli are encoded atstudy. This kind of encoding variability can arise, for example, ifparticipants’ attention fluctuates over the course of an experi-ment—some items will be encoded more strongly than others,leading to higher recall and familiarity scores at test.

An important property of encoding variability, which is not trueof sampling variability, is that it affects studied items and relatedlures in tandem. That is, encoding fluctuations that boost the

Figure 10. Plot of the probability that lures and studied items will triggermismatching recall, as a function of target–lure similarity. This probabilityis close to floor for studied items; the probability that lures will triggermismatching recall increases with increasing target–lure similarity.


memory signal triggered by a studied item also boost the memorysignal triggered by lures that are similar to that studied item (e.g.,if cat is encoded so as to be especially familiar, the related lure catswill also be highly familiar). In contrast, sampling variabilityoperates independently on each input feature; in small networkswhere sampling variability is the dominant source of variance,noise associated with sampling of nonshared (discriminative) fea-tures of overlapping stimuli counteracts much of the shared vari-ability in memory scores triggered by these items. We revisit thisissue later, when we explore how lure relatedness interacts withtest format (Simulation 4: Lure-Relatedness and Test-Format In-teractions, below).

Another source of variability in recall and familiarity scores isvariability in preexperimental exposure to stimuli: Some stimulihave been encountered extensively prior to the experiment, inmany different contexts; other stimuli are relatively novel; forevidence that preexperimental presentation frequency affects rec-ognition memory, see Dennis and Humphreys (2001). Variabilityin preexperimental exposure (like encoding variability, but unlikesampling variability) affects studied items and related lures intandem.

Finally, in addition to variability in how much test items overlapwith preexperimental memory traces, there is also variability inhow much items overlap with other items presented in the exper-iment; this kind of variability also affects studied items and relatedlures in tandem. Overlap-related variability is already present inthe model, but its effect on performance is typically dwarfed bysampling variability. Consequently, variability in overlap shouldplay a much larger role, proportionally, in larger networks withminimal sampling variability.

Sources of Variability: Summary

In summary, the basic model (as described above) is stronglyinfluenced by sampling variability and lacks other plausiblesources of variability such as encoding variability. Given thatsampling variability is not likely to be a factor in humanrecognition-memory performance, one might conclude that thissource of variability should be eliminated and other sources incor-porated. Unfortunately, this is not practical at present—modelsthat are large enough to eliminate sampling variability cannot befeasibly run on available computational hardware. Furthermore,adding more variability on top of sampling variability in our smallnetworks leads to poor performance unless other steps are taken tocompensate for increased variability (e.g., increasing the learningrate).

Because of these limitations, we refrain in this article frommaking strong predictions about how manipulations affect vari-ance. Nevertheless, we can still use the basic model to explainmany phenomena that do not depend on the exact source ofvariability. Also, it is relatively straightforward to supplement thebasic model with other forms of variability on an as-needed basis,and we do this to make some important points in Simulation 4:Lure-Relatedness and Test-Format Interactions, below.

Part 2: Applications to Behavioral Phenomena

The simulations in this part of the article build on the basicresults described earlier by applying the models to a wide range of

empirical recognition-memory phenomena (e.g., how does inter-ference affect recognition performance in the two models?).Whenever possible, we present data that speak to the model’spredictions.

Simulation 4: Lure-Relatedness and Test-FormatInteractions

As we showed in Figure 9, the model predicts that the hip-pocampus should outperform cortex on standard YN tests whereparticipants have to discriminate between studied items and relatedlures. In this section, we show how giving participants a forcedchoice between studied items and corresponding related luresbenefits performance in the cortical model but not the hippocampalmodel, thereby mitigating the hippocampal advantage.

FC Testing and Covariance

In an FC test, participants are simultaneously presented with astudied item and a lure and are asked to select the studied item. Thespecific version of this test that boosts cortical performance in-volves pairing studied items with corresponding related lures (i.e.,lures related to the paired studied item; for example, study rat, testrat vs. rats).

The central insight as to why this format improves corticalperformance with related lures is that, even though related lurestrigger strong feelings of familiarity (because they overlap with thestudied items), corresponding studied items are reliably morefamiliar. Because performance in an FC test is based on thedifference in familiarity between paired items, even small differ-ences can drive good performance, as long as they are reliable.

The reliability of the familiarity difference depends on wherevariability comes from in the model. As discussed in the previoussection, some kinds of variability (e.g., differences in encodingstrength and preexperimental exposure) necessarily affect studied-and related-lure familiarity in tandem, whereas other kinds ofvariability (e.g., sampling variability) do not. When the formerkind of variability predominates, the familiarity values of studieditems and corresponding lures are highly correlated, and therefore,their difference is reliable. When sampling variability predomi-nates, the studied-lure familiarity difference is somewhat lessreliable.

More formally, the beneficial effect of using an FC test dependson covariance in the familiarity scores triggered by studied itemsand corresponding related lures (Hintzman, 1988, 2001). Thevariance of the studied-lure familiarity difference is given by thefollowing equation:

Var�S � L� � Var�S� � Var�L� � 2 � Cov�S, L�, (2)

where S represents familiarity of studied items, L that of lures, andVar is variance and Cov covariance between S and L. Equation 2shows that increasing covariance reduces the variance of the S–Lfamiliarity difference, which in turn boosts FC performance.

FC simulations using the basic cortical model. Our first con-cern was to assess how much FC testing with correspondingrelated lures benefits performance in our basic cortical model.Toward this end, we ran simulations using a paradigm introducedby Hintzman (1988); in these simulations, we compared FC per-formance with corresponding related lures (i.e., study A, B; test A


vs. A�, B vs. B�, where A� and B� are lures related to A and B,respectively) to FC performance with noncorresponding lures(e.g., study A, B; test A vs. B�, B vs. A�). To the extent that thereis covariance between studied items and corresponding lures, thiswill benefit performance in the corresponding-lure condition rel-ative to the noncorresponding lures.

As shown in Figure 11, FC performance is higher with corre-sponding related lures than with noncorresponding lures—thisreplicates the empirical results obtained by Hintzman (1988) andshows that there is some covariance present in the basic corticalmodel. To quantify the level of covariance underlying these re-sults, we computed the following ratio:

R � �2 � Cov�S, L��/�Var�S� � Var�L��. (3)

When R � 1, covariance completely offsets studied and lurevariance, and the studied-lure familiarity difference is completelyreliable (i.e., variance � 0); R � 0 means that there is no covari-ance. For target–lure similarity � .92, the covariance ratio R � .27in the corresponding condition, and R � �.01 in the noncorre-sponding condition. Thus, the model exhibits roughly one third themaximal level of covariance possible.

In summary, although the basic model qualitatively exhibits anFC advantage with corresponding related lures, this advantage isnot quantitatively very large. This is because the dominant sourceof variability in the basic cortical model is sampling variability,which—as discussed above—does not reliably affect studied itemsand corresponding lures in tandem.

Cortical and hippocampal simulations with encoding variabil-ity. Next, we wanted to explore a more realistic scenario inwhich the contribution of sampling variability to overall variabilitywas small relative to other forms of variability (such as encodingvariability) that affect studied items and corresponding lures intandem. Increasing the relative contribution of encoding variabilityshould increase covariance and thereby increase the extent towhich the cortical model benefits from use of an FC-correspondingtest. We were also interested in how test format affects the hip-pocampal model’s ability to discriminate between studied itemsand related lures (when encoding variability is high). To addressthese issues, we ran simulations with added encoding variability inboth the cortical and hippocampal models where we manipulated

test format (FC corresponding vs. FC noncorresponding vs. YN)and lure relatedness.

Method. We added encoding variability using the following simplemanipulation: For each item at study, the learning rate was scaled by arandom number from the 0-to-1 uniform distribution. However, this ma-nipulation by itself did not achieve the desired result; the influence ofencoding variability was still too small relative to sampling variability, andoverall performance levels with added encoding variability were unaccept-ably low. To boost the relative impact of encoding variability (and overallperformance), we also increased the learning rate in both models to 3 timesits usual value. Under this regime, random scaling of the learning rate atstudy has a much larger effect on studied-item (and related-lure) familiaritythan random differences in how well features are sampled. We should notethat using a large learning rate has some undesirable side effects (e.g.,increased interference), but these side effects are orthogonal to the ques-tions being asked here. As with the hippocampal related-lure simulationspresented earlier, the hippocampal simulations presented here used a recall-to-reject decision rule. We applied this rule to FC testing in the followingmanner: If one item triggered mismatching recall but the other item did not,the second item was selected; otherwise, the item triggering a highermatch–mismatch recall score was selected.

Cortical FC results. As expected, the corresponding versusnoncorresponding difference for the cortical model is much largerwhen encoding variability is present (see Figure 12A) than whenencoding variability is absent (see Figure 11). Computing theaverage covariance/variance ratio for the .92 target–lure overlapcondition shows that R � .62 for corresponding lures versus R ��.05 for noncorresponding lures. This is more than double thecovariance in the basic model (.62 vs. .27) and confirms ourintuition that decreasing the contribution of sampling variabilityrelative to encoding variability would increase covariance andboost performance in the FC-corresponding condition.

Hippocampal FC results. In contrast to the cortical modelresults, FC-corresponding and FC-noncorresponding performanceare almost identical in the hippocampal model (see Figure 12A). Itseems clear that the same arguments about covariance benefitingFC-corresponding performance should hold for hippocampus aswell as for cortex. Why then does the hippocampus behave dif-ferently than the cortex in this situation? This can be explained bylooking at what happened on trials where the studied item was notrecalled—on these trials, participants can still respond correctly ifthe lure triggers mismatching recall (and is rejected on this basis).The key insight is that studied recall and lure misrecall are inde-pendent when noncorresponding lures are used (in effect, partici-pants get two independent chances to make a correct response), butthey are highly correlated when corresponding lures are used—ifthe studied item does not trigger any recall, the corresponding lureprobably will not trigger any recall either. Thus, using correspond-ing lures can actually hurt performance in the hippocampal modelby depriving participants of an extra, independent chance to re-spond correctly (via recall-to-reject) on trials where studied recallfails. This harmful effect of covariance cancels out the beneficialeffects of covariance described earlier.

Because the cortical model benefits from FC-corresponding (vs.noncorresponding) testing but the hippocampal model does not,the performance of the cortical model relative to the hippocampalmodel is better in this condition.

YN results. The results of our YN simulations with encodingvariability (Figure 12B) are identical to the results of our earlierYN–related-lure simulations. As before, we found that the hip-

Figure 11. Forced-choice (FC) accuracy in the cortical model as a func-tion of target–lure similarity, using corresponding and noncorrespondingFC testing. For high levels of target–lure similarity, FC performance isslightly better with corresponding lures than with noncorresponding lures.


pocampal model is much better than the cortical model at discrim-inating between studied items and related lures on YN tests.

Tests of the Model’s Predictions

One way to test the model’s predictions is to look at recognitionin patients with focal, relatively complete hippocampal damage.Presumably, these patients are relying exclusively on MTLC fa-miliarity when making recognition judgments (in contrast to con-trols, who have access to both hippocampal recall and MTLCfamiliarity). As such, patients should perform poorly relative tocontrols on tests where hippocampus outperforms cortex, and theyshould perform relatively well on tests where hippocampus andcortex are evenly matched. Applying this logic to the results shownin Figure 12, patients should be impaired on YN recognition testswith related lures, but they should perform relatively well onFC-corresponding tests with related lures and on tests with unre-lated lures (regardless of test format).

To test this prediction, we collaborated with Andrew Mayes andJuliet Holdstock to test patient Y.R., who suffered focal hippocam-pal damage sparing surrounding MTLC regions (for details of theetiology and extent of Y.R.’s lesion, see Holdstock et al., 2002).Y.R. is severely impaired at recalling specific details—thus, Y.R.has to rely almost exclusively on MTLC familiarity when makingrecognition judgments. Holdstock et al. (2002) developed YN andFC tests with highly related lures that were closely matched fordifficulty and administered these tests to patient Y.R. and hercontrols. Figure 13 shows sample stimuli from this experiment.Results from this experiment can be compared with results from 15other YN item-recognition tests and 25 other FC item-recognitiontests that used lures that were less strongly related to studied items(Mayes et al., 2002); we refer to these tests as the YN–low-relatedness and FC–low-relatedness tests, respectively.

Figure 14 shows that, exactly as we predicted, Y.R. was signif-icantly impaired on a YN recognition test that used highly relatedlures but showed relatively spared performance on an FC versionof the same test (Y.R. actually performed slightly better than thecontrol mean on this test). This pattern cannot be explained interms of difficulty confounds (i.e., Y.R. performing worse, relativeto controls, on the more difficult test)—controls found the YN testwith highly related lures to be slightly easier than the FC test.Figure 14 also shows that Y.R. was, on average, unimpaired onYN–low-relatedness and FC–low-relatedness tests. Y.R. per-formed worse on the YN test with highly related lures than on anyof the 15 YN–low-relatedness tests. This difference cannot beattributed to the YN–low-relatedness tests being easier than theYN test with highly related lures: Y.R. showed unimpaired per-formance on the 8 YN–low-relatedness tests that controls found tobe more difficult than the YN test with highly related lures; forthese 8 tests, her mean z score was 0.04 (SD � 0.49; minimum ��0.54; maximum � 0.65; J. Holdstock, personal communication

Figure 13. Sample stimuli from the Holdstock et al. (2002) related-lureexperiment. Participants studied pictures of objects (e.g., the horse shownin the upper left). At test, participants had to discriminate studied picturesfrom three highly related lures (e.g., the horses shown in the upper right,lower left, and lower right). From “Under What Conditions Is RecognitionSpared Relative to Recall After Selective Hippocampal Damage in Hu-mans?” by J. S. Holdstock et al., 2002, Hippocampus, 12, p. 344. Copyright2002 by Wiley. Reprinted with permission.

Figure 12. Cortical and hippocampal related-lure simulations incorporat-ing strong encoding variability. A: Results of forced-choice (FC) simula-tions. When encoding variability is present, the cortical model benefitsvery strongly from use of corresponding versus noncorresponding lures(more so than in our simulations without encoding variability). In contrast,the hippocampal model (using recall-to-reject) performs equally well withcorresponding and noncorresponding lures. B: Results of yes–no (YN)simulations with the same parameters. As in our previous related-luresimulations (see Figure 9), the cortical model is severely impaired relativeto the hippocampal model on these tests. Hippo � hippocampus; MTLC �medial temporal lobe cortex; C � corresponding lures; N � noncorre-sponding lures.


December 15, 2000). We have yet to test the model’s predictionregarding use of FC-corresponding versus FC-noncorrespondingtests with related lures; on the basis of the results shown inFigure 12, the model predicts that Y.R. will be more stronglyimpaired on FC tests with noncorresponding (vs. corresponding)related lures.

Simulation 5: Associative Recognition and Sensitivity toConjunctions

In this section, we explore the two networks’ performance onassociative-recognition tests. On these tests, participants have todiscriminate between studied pairs of stimuli (A–B, C–D) andassociative lures generated by recombining studied pairs (A–D,B–C). To show above-chance associative-recognition perfor-mance, a network must be sensitive to whether features occurredtogether at study; sensitivity to individual features does not helpdiscriminate between studied pairs and recombined lures. Thehippocampus’s ability to rapidly encode and store feature conjunc-tions is not in dispute—this is a central feature of practically alltheories of hippocampal functioning, including ours (see, e.g.,Rolls, 1989; Rudy & Sutherland, 1995; Squire, 1992b; Teyler &Discenna, 1986). In contrast, many theorists have argued thatneocortex is not capable of rapidly forming new conjunctiverepresentations (i.e., representations that support differential re-sponding to conjunctions vs. their constituent elements) on itsown; see O’Reilly and Rudy (2001) for a review.

Associative recognition can be viewed as a special case of therelated-lure paradigm described earlier. As such, we would expectthe hippocampus to outperform cortex on YN associative-recognition tests because of its superior pattern-separation abilitiesand its ability to reject similar lures based on mismatching recall.

Associative Recognition

Method. In our associative-recognition simulations, 20 item pairs werepresented at study—each pair consisted of a 12-slot pattern concatenatedwith another 12-slot pattern; at test, studied pairs were presented along withthree types of lures: associative (re-paired) lures, feature lures (generatedby pairing a studied item with a nonstudied item), and novel lures (gener-ated by pairing two nonstudied items). Our initial simulations used a YNtest format.

YN results. As expected, the hippocampal model outperformsthe cortical model on this YN associative-recognition test (seeFigure 15). We also found that cortical performance is well abovechance. This indicates that cortex is sensitive (to some degree) tofeature co-occurrence in addition to individual feature occurrence.

The ability of the cortical model to encode stimulus conjunc-tions can be explained in terms of the fact that cortex, like thehippocampus, uses sparse representations (as enforced by thekWTA algorithm). The kWTA algorithm forces units to compete torepresent input patterns, and units that are sensitive to multiplefeatures of a given input pattern (i.e., feature conjunctions) aremore likely to win the competition than units that are sensitiveonly to single input features. Representations are more conjunctivein the hippocampus than in cortex because representations aremore sparse (i.e., there is stronger inhibitory competition) in thehippocampus than in the cortex. For additional computationalsupport for the notion that cortex should encode low-order con-junctive representations, see O’Reilly and Busby (2002).

Effects of Test Format

In Simulation 4: Lure-Relatedness and Test-Format Interac-tions, above, we showed how giving the models a forced choicebetween studied items and corresponding related lures mitigatesthe hippocampal advantage for related lures. Analogously, givingthe models a forced choice between overlapping studied pairs andlures (FC-OLAP testing: study A–B, C–D; test A–B vs. A–D)mitigates the hippocampal advantage for associative recognition.In both cases, performance suffers because the hippocompal model

Figure 15. Results of yes–no (YN) associative-recognition simulations inthe cortical (MTLC) and hippocampal (Hippo) models. With parametersthat yield matched performance for unrelated (novel) lures, cortex isimpaired relative to the hippocampus at associative recognition; nonethe-less, cortex performs well above chance on the associative-recognitiontests.

Figure 14. Performance of patient Y.R. relative to controls on matchedyes–no (YN) and forced-choice (FC) corresponding tests with highlyrelated (High) lures; the graph also plots Y.R.’s average performance on 15YN tests and 25 FC tests with less strongly related (Low) lures. Y.R.’sscores are plotted in terms of number of standard deviations (SDs) aboveor below the control mean. For the YN and FC low-relatedness tests, errorbars indicate the maximum and minimum z scores achieved by Y.R.(across the 15 YN tests and the 25 FC tests, respectively). Y.R. wassignificantly impaired relative to controls on the YN test with highlyrelated lures (i.e., her score was � 1.96 SDs below the control mean), butY.R. performed slightly better than controls on the FC test with highlyrelated lures. Y.R. was not significantly impaired, on average, on the teststhat used less strongly related lures.


does not get an extra chance to respond correctly (via recall-to-reject) when studied recall fails.

Typically, FC-OLAP tests are structured in a way that empha-sizes the shared item: Participants are asked, “Which of these itemswas paired with A: B or D?” This encourages participants to use astrategy where they cue with the shared item (A) and select thechoice (B or D) that best matches retrieved information. Theproblem with this algorithm is that success or failure dependsentirely on whether the shared cue (A) triggers recall; if A fails totrigger recall of B, participants are forced to guess. In contrast, onFC associative-recognition tests with nonoverlapping choices (FC-NOLAP tests: study A–B, C–D, E–F; test A–B vs. C–F), partici-pants have multiple, independent chances to respond correctly;even if A does not trigger recall of B, participants can still respondcorrectly if they recall that C was paired with D (not F).

To demonstrate how recall-to-reject differentially benefits FC-NOLAP performance (relative to FC-OLAP performance), we ranFC-NOLAP and FC-OLAP simulations in the hippocampal modelusing a recall-to-reject rule. For comparison purposes, we rananother set of simulations where decisions were based purely onthe amount of matching recall.

Test Format Simulation

Method. These simulations used a cued recall algorithm where, foreach test pair (e.g., A–B), we cued with the first pair item and measuredhow well recalled information matched the second pair item. On FC-OLAPtests, we cued with the shared pair item (A) for both test alternatives (A–Bvs. A–D). On FC-NOLAP tests, we cued with the first item of each testalternative (A from A–B and C from C–D). We had to adjust some modelparameters to get the model to work well using partial cues; specifically,we used a higher than usual learning rate (.03 vs. .01) to help foster patterncompletion of information not in the cue, and we increased the activationcriterion for counting a feature as recalled (from .90 to .95) to compensatefor the fact that the output of the model was less clean with partial cues.

Results. As expected, FC-NOLAP performance is higher inthe recall-to-reject condition (vs. the match-only condition), butFC-OLAP performance does not benefit at all (see Figure 16).Because of this differential benefit, FC-NOLAP performance isbetter overall than FC-OLAP performance in the recall-to-rejectcondition. Consistent with this prediction, Clark, Hori, and Callan(1993) found better performance on an FC-NOLAP associative-recognition test than on an FC-OLAP associative-recognition test.They explained this finding in a manner that is consistent with ouraccount—they argued that participants were using recall of studiedpairs to reject lures and that participants had more unique (inde-pendent) chances to recall useful information in the FC-NOLAPcondition.


The implications of the above simulation results for patientperformance are clear: Patients with focal hippocampal lesionsshould be impaired, relative to controls, on YN associative-recognition tests, but they should be relatively less impaired onFC-OLAP associative-recognition tests.

No one has yet conducted a direct comparison of how wellpatients with hippocampal damage perform relative to controls asa function of test format. However, there are several relevant datapoints in the literature. Kroll, Knight, Metcalfe, Wolf, and Tulving

(1996) studied associative-recognition performance in a patientwith bilateral hippocampal damage (caused by anoxia), as well asin other patients with less focal lesions. In Experiment 1 of theKroll et al. study, participants studied two-syllable words (e.g.,barter, valley) and had to discriminate between studied words andwords created by recombining studied words (e.g., barley). Resultsfrom the patient with bilateral hippocampal damage, as well ascontrol data, are plotted in Figure 17. In keeping with the model’spredictions, the patient showed impaired YN associative-recognition performance, but YN discrimination with novel lures(where neither part of the stimulus was studied) was intact. Fur-thermore, even though the patient was impaired at associativerecognition, the patient’s performance in this experiment wasabove chance. This is consistent with the idea that cortex issensitive (to some degree) to feature conjunctions. However, thisstudy does not speak to whether cortex can form novel associationsbetween previously unrelated stimuli—because stimuli (includinglures) were familiar words, participants did not necessarily have toform a new conjunctive representation to solve this task.

Two studies (Mayes et al., 2001; Vargha-Khadem et al., 1997)have examined how well patients with focal hippocampal damageperform on FC-OLAP tests where participants were cued with onepair item and had to say which of two items was paired with thatitem at study. The Vargha-Khadem et al. (1997) study used unre-lated word pairs, nonword pairs, familiar-face pairs, andunfamiliar-face pairs as stimuli, and the Mayes et al. (2001) studyused unrelated word pairs as stimuli. In both of these studies, thehippocampally lesioned patients were unimpaired. This is consis-tent with the model’s prediction that patients should performrelatively well, compared with controls, on FC-OLAP tests. Fur-thermore, despite the patients’ having hippocampal lesions, theirexcellent performance on these tests, coupled with the fact that thetests used novel pairings, provides clear evidence that cortex iscapable of forming new conjunctive representations (that arestrong enough to support recognition, if not recall) after a single

Figure 16. Associative-recognition performance in the hippocampalmodel, as a function of recall decision rule (match only vs. recall-to-reject)and test format (forced-choice with overlapping choices [FC-OLAP] vs.forced-choice with nonoverlapping choices [FC-NOLAP]). FC-NOLAPperformance benefits from use of the recall-to-reject rule, but FC-OLAPperformance does not benefit at all. When the recall-to-reject rule is used,FC-NOLAP performance is better than FC-OLAP performance.


study exposure. One caveat is that, although MTLC appears ca-pable of supporting good associative-recognition performancewhen the to-be-associated stimuli are the same kind (e.g., words),it performs less well when the to-be-associated stimuli are ofdifferent kinds (e.g., objects and locations; Holdstock et al., 2002;Vargha-Khadem et al., 1997). Holdstock et al. (2002) argued thatMTLC familiarity cannot support object–location associative rec-ognition because object and location information do not convergefully in MTLC.

As a final note, although the studies discussed above foundpatterns of spared and impaired recognition performance afterfocal hippocampal damage that are consistent with the model’spredictions, it is important to keep in mind that many studies havefound an across-the-board impairment in declarative-memorytasks following hippocampal damage. For example, Stark andSquire (2002) found that patients with hippocampal damage wereimpaired to a roughly equal extent on tests with novel versusre-paired lures. We address the question of why some hippocampalpatients show across-the-board versus selective deficits in Simu-lation 8: Lesion Effects in the Combined Model, below.

Simulation 6: Interference and List Strength

We now turn to the fundamental issue of interference: How doesstudying an item affect recognition of other, previously studieditems? In this section, we first review general principles of inter-ference in networks like ours that incorporate Hebbian LTP andLTD. We then show how a list-strength interference manipulation(described in detail below) differentially affects discriminationbased on hippocampal recall versus MTLC familiarity.

General Principles of Interference in Our Models

At the most general level, interference occurs in our modelswhenever different input patterns have overlapping internal repre-

sentations. In this situation, studying a new pattern tunes theoverlapping units so they are more sensitive to the new pattern andless sensitive to the unique (discriminative) features of otherpatterns.

Figure 18 is a simple illustration of this tuning process. It showsa network with a single hidden unit that receives input from fiveinput units. Initially, the hidden unit is activated to a roughly equalextent by two different input patterns, A and B. Studying PatternA has two effects: Hebbian LTP increases weights to active inputfeatures, and Hebbian LTD decreases weights to inactive inputfeatures. These changes bolster the extent to which Pattern Aactivates the hidden unit. The effects of learning on responding toPattern B are more complex: LTP boosts weights to features thatare shared by Patterns A and B, but LTD reduces weights tofeatures that are unique to Pattern B.

If one trains a network of this type on a large number ofoverlapping patterns (e.g., several pictures of fish), the networkbecomes more and more sensitive to features that are shared acrossthe entire item set (e.g., the fact that all studied stimuli have fins)and less and less responsive to discriminative features of individ-ual stimuli (e.g., the fact that one fish has a large green stripeddorsal fin). In the long run, this latter effect is harmful to recog-nition performance—if the network’s sensitivity to the unique,discriminative features of studied items and lures hits floor, thenthe network is not able to respond differentially to studied itemsand lures at test. However, in the absence of floor effects, theextent to which recognition is harmed depends on the extent towhich interference differentially affects responding to studieditems and lures. In the next section, we explore this issue in thecontext of our two models.

List-Strength Simulations

We begin our exploration of interference by simulating how liststrength affects recognition in the two models; specifically, howdoes strengthening some list items affect recognition of other(nonstrengthened) list items (Ratcliff, Clark, & Shiffrin, 1990)?

Figure 17. Associative recognition in a patient with bilateral hippocam-pal damage (from Kroll, Knight, Metcalfe, Wolf, & Tulving, 1996, Exper-iment 1); d� scores for the patient and controls were computed based onaverage hit and false-alarm rates published in Table 3 of Kroll et al. Thepatient performed better than adult controls at discriminating studied itemsfrom novel lures but was worse than controls at discriminating studieditems from feature lures (where one part of the stimulus was old and onepart was new) and was much worse than controls when lures were gener-ated by recombining studied stimuli. The pattern reported here is qualita-tively consistent with the model’s predictions as shown in Figure 15,above. YN � yes–no.

Figure 18. Illustration of how Hebbian learning causes interference inour models. Initially (top two squares), the hidden unit responds equally toPatterns A and B. The effects of studying Pattern A are shown in thebottom two squares. Studying Pattern A boosts weights to features that arepart of Pattern A (including features that are shared with Pattern B) andreduces weights to features that are not part of Pattern A. These changesresult in a net increase in responding to Pattern A and a net decrease inresponding to Pattern B. LTP � long-term potentiation; LTD � long-termdepression.


Method. In our list-strength simulations, we compared two conditions:a weak-interference condition and a strong-interference condition. In theweak-interference condition, the model was given a study list consisting oftarget items presented once and interference items presented once. Thestrong-interference condition was the same except that interference itemswere strengthened at study by presenting them multiple times. In bothconditions, the model had to discriminate between target items and non-studied lures at test. If strengthening of interference items (in the strong-interference condition) impairs target versus lure discrimination relative tothe weak-interference condition, this is an LSE. A simple diagram of theprocedure is provided in Table 1.

The study list was comprised of 10 target items followed by 10 inter-ference items. Interference-item strength was manipulated by increasingthe learning rate for these items (from .01 to .02). In our models, strength-ening by repetition and strengthening by increasing the learning rate havequalitatively similar effects; however, quantitatively, repetition has a largereffect on weights (e.g., doubling the number of presentations leads to moreweight change than doubling the learning rate) because the initial studypresentation alters how active units are on the next presentation, andgreater activity leads to greater learning (according to the Hebb rule). Wealso manipulated average between-item overlap (ranging from 10% to50%) to see how this factor interacts with list strength—intuitively, in-creasing overlap should increase interference.

Results. In the cortical network (see Figure 19A), there is noeffect of list strength on recognition when input-pattern overlap isrelatively low (up to .26), but the LSE is significant for higherlevels of input overlap. In contrast, the hippocampal networkshows a significant LSE for all levels of input overlap (see Figure19B); the size of the hippocampal LSE increases with increasingoverlap (except in the .5 overlap condition, where the LSE iscompressed by floor effects). Figure 19C directly compares thesize of the LSE in the two models.

We also measured the direct effect of strengthening interferenceitems on memory for those items; both models exhibit a robustitem-strength effect whereby memory for interference items isbetter in the strong-interference condition (e.g., for 20% inputoverlap, interference-item d� increases from 2.13 to 3.22 in thehippocampal model; in the cortical model, d� increases from 2.08to 3.12), thereby confirming that our strengthening manipulation iseffective.

The data are puzzling: For moderate amounts of overlap, thehippocampus shows an interference effect despite its ability tocarry out pattern separation, and cortex—which has higher base-line levels of pattern overlap—does not show an interferenceeffect. We address the hippocampal results first.

Interference in the Hippocampal Model

Understanding the hippocampal LSE is quite straightforward.Even though there is less overlap between representations in thehippocampus than in cortex, there is still some overlap (primarilyin CA1, but also in CA3). These overlapping units cause interfer-ence—specifically, recall of discriminative features of studieditems is impaired through Hebbian LTD occurring in the CA3–CA1 projection and (to a lesser extent) in projections coming intoCA3. Importantly, for low to moderate levels of input-patternoverlap, the amount of recall triggered by lures is at floor, andtherefore, it cannot decrease as a function of interference. Puttingthese two points together, the net effect of interference is to movethe studied distribution downward toward the at-floor lure distri-bution, which increases the overlap between distributions andtherefore impairs discriminability (see Figure 20).

Interference in the Cortical Model

Next, we need to explain why an LSE is not obtained in thecortical model (for low to moderate levels of input overlap). Thecritical difference between the cortical and hippocampal models is

Table 1List-Strength Procedure

Target items Interference items

Weak interference

bike robot apple cat tree towel

Strong interference

bike robot apple cat tree towelcat tree towelcat tree towel

Note. List-strength simulations compared weak-interference lists withstrong-interference lists. In both conditions, the model had to discriminatebetween targets (e.g., bike) and nonstudied lures (e.g., coin) at test.

Figure 19. Results of list-strength simulations in the two models. A:Effect of list strength on recognition in medial temporal lobe cortex(MTLC). B: Effect of list strength on recognition in the hippocampus(Hippo). C: Size of the list-strength effect (LSE) in MTLC and thehippocampus; this panel replots data from A and B as list-strength differ-ence scores (weak-interference d� � strong-interference d�) to facilitatecomparison across models. For low to moderate levels of overlap (up to.26), there is a significant LSE in the hippocampal model but not in thecortical model; for higher levels of overlap, there is an LSE in both models.YN � yes–no; Strong int. � strong interference; Weak int. � weakinterference.


that lure familiarity is not at floor in the cortical network, therebyopening up the possibility that lure familiarity (as well as studiedfamiliarity) might actually decrease as a function of interference.Discriminability is a function of the difference in studied and lurefamiliarity (as well as the variance of these distributions); there-fore, if lure familiarity decreases as much as (or more) than studiedfamiliarity as a function of interference, overall discriminabilitymay be unaffected. This is in fact what occurs in the corticalmodel.

Figure 21A shows how raw familiarity scores triggered bytargets and lures change as a function of interference (for 20%overlap). Initially, both studied and lure familiarity decrease; thisoccurs because interference reduces weights to discriminative fea-tures of both studied items and lures. There is also an interactionwhereby lure familiarty decreases more quickly than studied fa-miliarity, so the studied–lure gap in familiarity increases slightly(see Figure 21B; note that although the increase is numericallysmall, it is highly significant).

However, with more interference, the studied–lure gap in famil-iarity starts to decrease. This occurs because weights to discrimi-native features of lures eventually approach floor (and—as such—cannot show any additional decrease as a function of interference).Also, raw familiarity scores start to increase; this occurs because,in addition to reducing weights to discriminative features, inter-ference also boosts weights to shared-item features. At first, theformer effect outweighs the latter, and there is a net decrease infamiliarity. However, when weights to discriminative featuresapproach floor, these weights no longer decrease enough to offsetthe increase in weights to shared features, and there is a netincrease in familiarity.

For these parameters, list strength does not substantially affectthe variance of the familiarity signal; variance increases numeri-cally with increasing list strength, but the increase is extremelysmall (e.g., 10-times strengthening of the 10 interference itemsleads to a 3% increase in variance). However, we cannot rule outthe possibility that adding other forms of variability to the model(and eliminating sampling variability; see the Sources of Variabil-ity section above) might alter the model’s predictions about howlist strength affects variance.

Differentiation. The finding that lure familiarity initially de-creases faster than studied familiarity can be explained in terms ofthe principle of differentiation, which was first articulated byShiffrin et al. (1990); see also McClelland and Chappell (1998).Shiffrin et al. argued that studying an item makes its memoryrepresentation more selective, such that the representation is lesslikely to be activated by other items.

In our model, differentiation is a simple consequence of Heb-bian learning (as it is in McClelland & Chappell, 1998). Asdiscussed above, Hebbian learning tunes MTLC units so that theyare more sensitive to the studied input pattern (because of LTP)and less sensitive to other, dissimilar input patterns (because ofLTD). Because of this LTD effect, studied-item representations areless likely to be activated by interference items than lure-itemrepresentations; as such, studied items suffer less interference thanlures. As an example of how studying an item pulls its represen-tation away from other items, with 20% input overlap, the averageamount of MTLC overlap between studied target items and inter-ference items (expressed in terms of vector dot product) is .150,whereas the average overlap between lures and interference itemsis .154; this difference is highly significant.

Boundary conditions on the null LSE. It should be clear fromthe above explanation that we do not always expect a null LSE for

Figure 20. Studied-recall histograms for the strong- and weak-interference conditions, 20% overlap condition. Increasing list strengthpushes the studied-recall distribution to the left (toward zero).

Figure 21. A: Plot of how target and lure familiarity are affected by liststrength (with 20% input overlap). Initially, target and lure familiaritydecrease; however, with enough interference, target and lure familiaritystart to increase. B: Plot of the difference in target and lure familiarity, asa function of list strength; initially, the difference increases slightly, butthen it decreases. Lrate � learning rate.


MTLC familiarity. With enough interference, the cortical model’soverall sensitivity to discriminative features always approachesfloor, and the studied and lure familiarity distributions converge.The amount of overlap between items determines how quickly thenetwork arrives at this degenerate state—more overlap yieldsfaster degeneration. When overlap is high, raw familiarity scoresincrease (and the familiarity gap decreases) right from the start;this is illustrated in Figure 22, which plots target and lure famil-iarity as a function of list strength, for 40.5% input overlap.


The main novel prediction from our models is that (modulo theboundary conditions outlined above) recognition based on hip-pocampal recall should exhibit an LSE, whereas recognition basedon MTLC familiarity should not.

Consistent with the hippocampal model’s prediction, some stud-ies have found an LSE for cued recall (see, e.g., Kahana, Rizzuto,& Schneider, 2003; Ratcliff et al., 1990), although not all studiesthat have looked for a cued-recall LSE have found one (see, e.g.,Bauml, 1997). However, practically all published studies that havelooked for an LSE for recognition have failed to find one (Mur-nane & Shiffrin, 1991a, 1991b; Ratcliff et al., 1990; Ratcliff, Sheu,& Gronlund, 1992; Shiffrin, Huber, & Marinelli, 1995; Yonelinas,Hockley, & Murdock, 1992). Although this null LSE for recogni-tion is consistent with our cortical model’s predictions (viewed inisolation), it is nevertheless somewhat surprising that overall rec-ognition scores do not reflect the hippocampal model’s tendency toproduce a recognition LSE.

One way to reconcile the null LSE for recognition with themodel’s predictions is to argue that hippocampal recall was notmaking enough of a contribution, relative to MTLC familiarity, onexisting tests. This explanation leads to a clear prediction: LSEsshould be obtained for recognition tests and measures that loadmore heavily on the recall process. Norman (2002), tested thisprediction in two distinct ways.

Self-report measures. In one experiment, Norman (2002) col-lected self-report measures of recall and familiarity—whenever a

participant called an item old, he or she was asked whether he orshe specifically remembered details from when the item wasstudied or whether the item just seemed familiar. To estimaterecall-based discrimination, Norman plugged the probability ofsaying “remember” to studied items and lures into the formula ford� and computed familiarity-based discrimination using the inde-pendence remember–know technique described in Jacoby, Yoneli-nas, and Jennings (1997).

Norman (2002) found that the effect of list strength on old/newrecognition sensitivity was nonsignificant in this experiment, rep-licating the null LSE obtained by Ratcliff et al. (1990). However,if one breaks recognition into its component processes, it is clearthat list strength does affect performance (see Figure 23). Aspredicted, there was a significant LSE for discrimination based onrecall; in contrast, list strength had no effect whatsoever onfamiliarity-based discrimination.2 We should emphasize that thetechnique we used to estimate familiarity-based discriminationassumes that recall and familiarity are independent–the indepen-dence assumption is discussed in more detail in Simulation 7: TheCombined Model and the Independence Assumption, below. Also,as discussed by Norman, one cannot conclusively rule out (in thiscase) the possibility that the observed LSE for recall-based dis-crimination was caused by shifting response bias, with no realchange in sensitivity. Nonetheless, the overall pattern of results ishighly consistent with the predictions of our model.

Lure relatedness. Norman (2002), also looked at how liststrength affected discrimination in the plurals paradigm (Hintzmanet al., 1992), where participants have to discriminate betweenstudied words, related switched-plurality (SP) lures (e.g., studyscorpion, test scorpions), and unrelated lures. The model predictsthat the ability to discriminate between studied words and relatedSP lures should depend on recall (see Simulation 3: YN Related-Lure Simulations, above). Thus, we should find a significant LSEfor studied versus SP discrimination but not necessarily for studied

2 Norman (2002) did not report how list strength affects familiarity-based discrimination—these results are being presented for the first timehere.

Figure 22. Plot of how target and lure familiarity are affected by liststrength with 40.5% input overlap. When overlap is high, target and lurefamiliarity increase right from the start, and the target–lure familiarity gapmonotonically decreases. Lrate � learning rate.

Figure 23. Plot of the size of the list-strength effect (LSE) for threedependent measures: d�(R), recall-based discrimination; d�(F), familiarity-based discrimination; and d�(Old), discrimination computed based on old/new responses. Error bars indicate 95% confidence intervals. The LSE wassignificant for d�(R) but not for d�(F) or d�(Old). Strong int. � stronginterference; Weak int. � weak interference.


versus unrelated discrimination, which can also be supported byfamiliarity.

Furthermore, we can also look at SP versus unrelated pseudo-discrimination, that is, how much more likely participants are tosay old to related versus unrelated lures. Familiarity boostspseudodiscrimination (insofar as SP lures are more familiar thanunrelated lures), but recall of plurality information lowers pseudo-discrimination (by allowing participants to confidently reject SPlures). Hence, if increasing list strength reduces recall of pluralityinformation (but has no effect on familiarity-based discrimination),the net effect should be an increase in pseudodiscrimination (anegative LSE).

As predicted, the LSE for studied versus SP-lure discriminationis significant, the LSE for studied versus unrelated-lure discrimi-nation is nonsignificant, and there is a significant negative LSE forSP-lure versus unrelated-lure pseudodiscrimination (see Figure24).

In summary, we obtained data consistent with the model’sprediction of an LSE for recall-based discrimination and with anull LSE for familiarity-based discrimination, using two verydifferent methods of isolating the contributions of these processes(collecting self-report data and using related lures). The model’spredictions regarding the boundary conditions of the null LSE forfamiliarity-based discrimination remain to be tested. For example,our claim that discrimination asymptotically goes to zero withincreasing list strength implies that it should be possible to obtainan LSE (in paradigms that have previously yielded a null LSE) byincreasing the number of training trials for interference items.

List-Length Effects

Thus far, our interference simulations have focused on liststrength. Here, we briefly discuss how list length affects perfor-mance in the two models. In contrast to the list-strength paradigm,

which measures how strengthening items already on the list inter-feres with memory for other items, the list-length paradigm mea-sures how adding new (previously nonstudied) items to the listinterferes with memory for other items.

The basic finding is that our model, as currently configured,makes the same predictions regarding list-length and list-strengtheffects—adding new items and strengthening already-presenteditems induce a comparable amount of weight change and thereforeresult in comparable levels of interference. As with list strength,the model predicts a dissociation whereby (so long as overlapbetween items is not too high) list length affects discriminationbased on hippocampal recall but not MTLC familiarity. Yonelinas(1994) obtained evidence consistent with this prediction using theprocess-dissociation procedure (which assumes independence).

However, some extant evidence appears to contradict the mod-el’s prediction of parallel effects of list length and list strength. Inparticular, Murnane and Shiffrin (1991a) and others have obtaineda dissociation whereby list length impairs recognition sensitivitybut a closely matched list-strength manipulation does not. Thisfinding implies that adding new items to the list is more disruptiveto existing weight patterns than repeating already-studied items.We are currently exploring different ways of implementing thisdynamic in our model.

One particularly promising approach is to add a large, transientfast-weight component to the model that reaches its maximumvalue in one study trial and then decays exponentially; subsequentpresentations of the same item simply reset fast weights to theirmaximum value, and decay begins again (see Hinton & Plaut,1987, for an early implementation of a similar mechanism). Thisdynamic is consistent with neurobiological evidence showing alarge, transient component to LTP (see, e.g., Bliss & Collingridge,1993; Malenka & Nicoll, 1993).

The assumptions outlined above imply that the magnitude offast weights at test (for a particular item) is a function of the timeelapsed from the most recent presentation of the item; the numberof times that the item has been repeated before this final presen-tation does not matter. As such, presenting interference items forthe first time (increasing list length) should have a large effect onfast weights, but repeating already-studied interference items (in-creasing list strength) should have less of an effect on the config-uration of fast weights at test. Preliminary cortical-model simula-tions incorporating fast weights (in addition to standard,nondecaying weights) have shown a list-length–list-strength dis-sociation, but more work is needed to explore the relative merits ofdifferent implementations of fast weights (e.g., should the fast-weight component incorporate LTD as well as LTP?) and how thepresence of fast weights interacts with the other predictions out-lined in this article.

The idea that list-length effects are attributable to quickly de-caying weights implies that interposing a delay between study andtest (thereby allowing the fast weights to decay) should greatlydiminish the list-length effect. In keeping with this prediction, arecent study with a 5-min delay between study and test did notshow a list-length effect (Dennis & Humphreys, 2001). The nextstep in testing this hypothesis will be to run experiments thatparametrically manipulate study–test lag while measuring list-length effects.

Figure 24. Results from the plurals list-strength effect (LSE) experiment.In this experiment, recognition sensitivity was measured using Az (an indexof the area under the receiver-operating characteristic curve; Macmillan &Creelman, 1991). The graph plots the size of the LSE for three differentkinds of discrimination: Studied (S) versus related switched-plurality (SP)lures; S versus unrelated (U) lures; and SP versus U lure pseudodiscrimi-nation. Error bars indicate 95% confidence intervals. There was a signif-icant LSE for S versus SP lure discrimination, and there was a significantnegative LSE for SP versus U lure pseudodiscrimination. Strong int. �strong interference; Weak int. � weak interference.


Simulation 7: The Combined Model and the IndependenceAssumption

Up to this point, we have explored the properties of hippocam-pal recall and MTLC familiarity by presenting input patterns toseparate hippocampal and neocortical networks—this approach isuseful for analytically mapping out how the two networks respondto different inputs, but it does not allow us to explore interactionsbetween the two networks. One important question that cannot beaddressed using separate networks is the statistical relationshipbetween recall and familiarity. As mentioned several timesthroughout this article, all extant techniques for measuring thedistinct contributions of recall and familiarity to recognition per-formance assume that they are stochastically independent (see,e.g., Jacoby et al., 1997). This assumption cannot be directly testedusing behavioral data because of the chicken-and-egg problemsdescribed in the introductory section, above.

To assess the validity of the independence assumption, weimplemented a combined model in which the cortical systemserves as the input to the hippocampus—this arrangement moreaccurately reflects how the two systems are connected in the brain.Using the combined model, we show here that there is no simpleanswer regarding whether or not hippocampal recall and MTLCfamiliarity are independent. Rather, the extent to which theseprocesses are correlated is a function of different (situationallyvarying) factors, some of which boost the correlation and some ofwhich reduce the correlation. In this section, we briefly describethe architecture of the combined model, and then we use the modelto explore two manipulations that push the recall–familiarity cor-relation in opposite directions: encoding variability and interfer-ence (list length).

Combined-Model Architecture

The combined model is structurally identical to the hippocampalmodel except that the projection from input to EC in has modifiableconnections (and 25% random connectivity) instead of fixed one-to-one connectivity. Thus, the input-to-EC in part of the combinedmodel has the same basic architecture and connectivity as the separatecortical model. This makes it possible to read out our act win famil-iarity measure from the EC in layer of the combined model (i.e., theEC in layer of the combined model serves the same functional role asthe MTLC layer of the separate cortical network).

There are, however, a few small differences between the corticalpart of the combined model and the separate cortical network.First, the EC in layer of the combined model is constrained tolearn slotted representations where only one unit in each 10-unitslot is strongly active; limiting the range of possible EC represen-tations makes it easier for the hippocampus to learn a stablemapping between CA1 representations and EC representations.Second, the EC in layer for the combined model has only 240units, compared with 1,920 units in the MTLC layer of the separatecortical network. This reduced size derives from computationalnecessity—use of a larger EC in would require a larger CA1,which together would make the simulations run too slowly oncurrent hardware. This smaller hidden layer in the combined modelmakes the familiarity signal more subject to sampling variability,and thus, recognition d� is somewhat worse, but otherwise itfunctions just as before. We used the same basic cortical and

hippocampal parameters as in our separate-network simulations,except that we used input patterns with 32.5% overlap—this levelof input overlap yields approximately 24% overlap between EC inpatterns at study.

In the combined model, the absolute size of the recall–familiarity correlation is likely to be inflated, relative to the brain,because of the high level of sampling variability present in ourmodel. Sampling variability leads to random fluctuations in thesharpness of cortical representations, which induce a correlation(insofar as sharper representations trigger larger familiarity scores,and they also propagate better into the hippocampus, bolsteringrecall). Because of this issue, our simulations below focus onidentifying manipulations that affect the size of the correlation,rather than characterizing the absolute size of the correlation.

Effects of Encoding Variability

Curran and Hintzman (1995) pointed out that encoding variabil-ity can boost the recall–familiarity correlation; if items vary sub-stantially in how well they are encoded, poorly encoded items willbe unfamiliar and will not be recalled, whereas well-encoded itemswill be more familiar and more likely to trigger recall. We ransimulations in the combined model where we manipulated encod-ing variability by varying the probability of partial encoding fail-ure (i.e., encoding only half of an item’s features) from 0 to .50.For each simulated participant, we read out the MTLC familiaritysignal (act win, read out from the EC in layer) and the hippocam-pal recall signal (match–mismatch between EC out and EC in) oneach trial and measured the correlation between these signals(across trials). The results of these simulations are presented inFigure 25; as expected, increasing encoding variability increasedthe recall–familiarity correlation in the model.

Interference-Induced Decorrelation

In the next set of simulations, we show how the presence ofinterference between memory traces can reduce the recall–familiarity correlation. In Simulation 6: Interference and ListStrength, above, we discussed how the two systems are

Figure 25. Simulations exploring the effect of encoding variability on therecall–familiarity correlation for studied items. As encoding variability(operationalized as the probability that the model will “blink” and fail toencode half of an item’s features at study) increases, the recall–familiaritycorrelation increases.


differentially affected by interference: Hippocampal recallscores for studied items tend to decrease with interference;familiarity scores decrease less, and sometimes increase, be-cause increased sensitivity to shared prototype features com-pensates for lost sensitivity to discriminative item-specific fea-tures. Insofar as items vary in how much interference they aresubject to (due to random differences in between-items overlap)and interference pushes recall and familiarity in different di-rections, it should be possible to use interference as a wedge todecorrelate recall and familiarity.

We ran simulations measuring how the recall–familiarity corre-lation changed as a function of interference (operationalized usinga list-length manipulation). There were 10 target items followed by0 to 150 (nontested) interference items. As expected, increasinglist length lowers the recall–familiarity correlation for studieditems (see Figure 26A) in the model.

We can get further insight into these results by looking at howinterference affects raw familiarity and recall scores for studieditems in these simulations (see Figure 26B). With increasinginterference, recall decreases sharply, but familiarity stays rela-tively constant; this differential effect of interference works todecorrelate the two signals. Once recall approaches floor, recalland familiarity are affected in a basically similar manner (i.e., notmuch); this lack of a differential effect explains why the recall–familiarity correlation does not continue to decrease all the way tozero.

In summary, the combined model gives us a principled means ofpredicting how different factors affect the statistical relationshipbetween recall and familiarity. Given the tight coupling of thecortical and hippocampal networks in the combined model, onemight think that a positive correlation is inevitable. However, theresults presented here show that—to a first approximation—inde-pendence can be achieved so long as factors that reduce thecorrelation (e.g., interference) exert more of an influence thanfactors that bolster the correlation (e.g., encoding variability).Future research will focus on identifying additional factors thataffect the size of the correlation.

Simulation 8: Lesion Effects in the Combined Model

In this section, we show how the combined model can providea more sophisticated understanding of the effects of different kindsof medial temporal lesions. Specifically, we show that (in thecombined model) partial hippocampal lesions can sometimes re-sult in worse overall recognition performance than complete hip-pocampal lesions. In contrast, increasing MTLC lesion size mono-tonically reduces overall recognition performance.

Partial Lesion Simulation

Method. We ran one set of simulations exploring the effects of focalhippocampal lesions and another set of simulations exploring the effects offocal MTLC lesions. In all of the lesion simulations, the size of the lesion(in terms of percentage of units removed) was varied from 0% to 95% in5% increments. In the hippocampal lesion simulations, we lesioned all ofthe hippocampal subregions (DG, CA1, CA3) equally by percentage; in theMTLC lesion simulations, we lesioned EC in. To establish comparablebaseline (prelesion) recognition performance between the hippocampal andcortical networks, we boosted the cortical learning rate to .012 instead of.004; this increase compensates for the high amount of sampling variabilitypresent in the (240-unit) EC in layer of the combined model. In thesesimulations, we used FC testing to maximize comparability with animallesion studies that had used this format (see, e.g., Baxter & Murray,2001b).

To simulate overall recognition performance when both processes arecontributing, we used a decision rule whereby if one item triggered a largerpositive recall score (match � mismatch) than the other, then that item wasselected; otherwise, if match � mismatch is less than or equal to 0, thedecision fell back on familiarity. This rule incorporates the assumption(shared by other dual-process models; e.g., Jacoby et al., 1997) that recalltakes precedence over familiarity. This differential weighting of recall canbe justified in terms of our finding that, in normal circumstances, hip-pocampal recall in the CLS model is more diagnostic than MTLC famil-iarity. Furthermore, it is supported by data showing that recall is associatedwith higher average recognition confidence ratings than familiarity (see,e.g., Yonelinas, 2001).

Hippocampal lesion results. Figure 27 shows how hippocam-pal FC performance, cortical FC performance, and combined FC

Figure 26. Simulations exploring how interference affects the recall–familiarity correlation. A: Increasing listlength reduces the recall–familiarity correlation for studied items. B: Increasing list length leads to a decreasein studied-item recall scores, but studied-item familiarity scores stay relatively constant.


performance (using the decision rule just described) vary as afunction of hippocampal lesion size. As one might expect, hip-pocampal FC performance decreases steadily as a function ofhippocampal lesion size, whereas cortical performance is unaf-fected by hippocampal damage (because familiarity is computedbefore activity is fed into the hippocampus). The most interestingresult is that hippocampal lesion size has a nonmonotonic effect oncombined recognition performance. At first, combined FC accu-racy decreases with increasing hippocampal lesion size; however,going from a 75% to a 100% hippocampal lesion actually im-proves combined FC performance.

Why is recognition performance worse for partial (75%) versuscomplete lesions? The key to understanding this finding is thatpartial hippocampal damage impairs the hippocampus’s ability tocarry out pattern separation. We assume that lesioning a layerlowers the total number of units but does not decrease the numberof active units; accordingly, representations become less sparse,and the average amount of overlap between patterns increases.There is neurobiological support for this assumption: In the brain,the activity of excitatory pyramidal neurons is regulated primarilyby inhibitory interneurons (Douglas & Martin, 1990); assumingthat both excitatory and inhibitory neurons are damaged by lesions,this loss of inhibition is likely to result in a proportional increasein activity for the remaining excitatory neurons.

Pattern-separation failure (induced by partial damage) results ina sharp increase in the amount of recall triggered by lures. Com-bined recognition performance in these participants suffers be-cause the noisy recall signal drowns out useful information that ispresent in the familiarity signal. Moving from a partial hippocam-pal lesion to a complete lesion improves performance by removingthis source of noise.

Figure 28 provides direct support for the noisy-hippocampustheory; it shows how the probability that lures will trigger apositive match � mismatch score increases steadily with increas-ing hippocampal lesion size until it reaches a peak of .41 (for 55%

hippocampal damage). However, as lesion size approaches 60%,the probability that lures and studied items will trigger a positivematch � mismatch score starts to decrease. This occurs for tworeasons: First, because of pattern-separation failure in CA3, allitems start to trigger recall of the same prototypical features (whichmismatch item-specific features of the test probe); second, CA1damage reduces the hippocampus’s ability to translate CA3 activ-ity into the target EC pattern. As the number of items triggeringpositive match–mismatch scores decreases, control of the recog-nition decision reverts to familiarity. This benefits recognitionperformance insofar as familiarity does a better job of discrimi-nating between studied items and lures than the recall signalgenerated by the lesioned hippocampus.

MTLC lesion results. Turning to the effects of MTLC lesions,we found that lesioning EC in hurts both cortical and hippocampalrecognition performance (see Figure 29). Therefore, combinedrecognition performance decreases steadily as a function of MTLClesion size. The observed deficits result from the fact that overlapbetween EC in patterns increases with lesion size—this has adirect adverse effect on cortically based discrimination; further-more, because EC in serves as the input layer for the hippocam-pus, increased EC in overlap leads to increased hippocampaloverlap, which hurts recall.

Relevant Data and Implications

The simulation results reported above are consistent with theresults of a recent meta-analysis conducted by Baxter and Murray(2001b). This meta-analysis incorporated results from several stud-ies that have looked at hippocampal and perirhinal (MTLC) lesioneffects on recognition in monkeys using a delayed-nonmatching-to-sample paradigm. Baxter and Murray found, in keeping withour results, that partial hippocampal lesions can lead to largerrecognition deficits than more complete lesions—in the meta-analysis, hippocampal lesion size and recognition impairment werenegatively correlated. The Baxter and Murray meta-analysis alsofound that perirhinal lesion size and recognition impairment were

Figure 27. Effect of hippocampal damage on forced-choice (FC) recog-nition performance. This graph plots FC accuracy based on medial tem-poral lobe cortex (MTLC) familiarity, hippocampal recall, and a combi-nation of the two signals, as a function of hippocampal lesion size. FCaccuracy based on MTLC familiarity is unaffected by hippocampal lesionsize. FC accuracy based on hippocampal recall declines steadily as lesionsize increases. FC accuracy based on a combination of recall and famil-iarity is affected in a nonmonotonic fashion by lesion size: Initially,combined FC accuracy declines; however, going from a 75% to a 100%hippocampal lesion leads to an increase in combined FC performance.

Figure 28. Plot of the probability that studied items and lures will triggera positive match � mismatch score, as a function of hippocampal lesionsize. For studied items, the probability declines monotonically as a functionof lesion size. For lures, the probability first increases, then decreases, asa function of lesion size.


positively correlated, just as we found that MTLC lesion size andimpairment were correlated in our simulations.

Baxter and Murray’s (2001b) results are highly controversial;Zola and Squire (2001) reanalyzed the data in the Baxter andMurray meta-analysis using a different set of statistical techniquesthat control, for example, for differences in mean lesion size acrossstudies and found that the negative correlation between hippocam-pal lesion size and impairment reported by Baxter and Murray wasno longer significant (although there was still a nonsignificanttrend in this direction; see Baxter & Murray, 2001a, for furtherdiscussion of this issue). Our results contribute to this debate byproviding a novel and principled account of how a negative cor-relation might come into being in terms of pattern-separationfailure resulting in a noisy recall signal that participants neverthe-less relied on when making recognition judgments. Of course, ouraccount is not the only way of explaining why partial hippocampallesions might impair recognition more than complete lesions.Another possibility, suggested by Mumby et al. (1996), is that adamaged hippocampus might disrupt neocortical processing viaepileptiform activity.

The results reported here speak to the debate over why hip-pocampally lesioned patients sometimes show relatively sparedrecognition performance on standard recognition tests with unre-lated lures and sometimes do not. The model predicts that patientswith relatively complete hippocampal lesions (that nonethelessspare MTLC) should show relatively spared performance andpatients with smaller hippocampal lesions (that reduce the diag-nosticity of recall without eliminating the signal outright) shouldshow an across-the-board declarative memory deficit without anyparticular sparing of recognition. This view implies that the rangeof etiologies that produce selective sparing should be quite narrow:If the lesion is too small, one ends up with a partially lesionedhippocampus that injects noise into the recognition process; if thelesion is too large, then one hits perirhinal cortex (in addition to thehippocampus), and this leads to deficits in familiarity-basedrecognition.

General Discussion

In this section of the article, we review how our modeling workspeaks to extant debates over the characterization of the respectivecontributions of recall (vs. familiarity) and hippocampus (vs.MTLC) to recognition. Next, we compare our model to otherneurally inspired and abstract computational models of recogni-tion. We conclude by discussing limitations of the models andfuture directions for research.

Implications for Theories of How Recall Contributes toRecognition

As discussed in the introductory section, above, the dual-processapproach to recognition has become increasingly prevalent inrecent years, but this enterprise is based on a weak foundation.Yonelinas, Jacoby, and their colleagues (see Yonelinas, 2001)have developed several techniques for measuring the contributionsof recall and familiarity based on behavioral data, but these tech-niques rely on a core set of assumptions that have not been tested;furthermore, some of these assumptions (e.g., independence) can-not be tested based on behavioral data alone because of chicken-and-egg problems.

More specifically, measurement techniques such as processdissociation and ROC analysis are built around dual-processsignal-detection theory (Jacoby et al., 1997; Yonelinas, 2001). Inaddition to the independence assumption, this theory assumes thatfamiliarity is a Gaussian signal-detection process but recall is adual high-threshold process (i.e., recall is all or none; studied itemsare sometimes recalled as being old, but lures never are; lures aresometimes recalled as being new, but studied items never are). TheCLS model does not assume any of the above claims to betrue—rather, its core assumptions are based on the functionalproperties of hippocampus versus MTLC. As such, we can use theCLS model to evaluate the validity of dual-process signal-detection theory.

Results from our cortical model are generally consistent with theidea that familiarity is a Gaussian signal-detection process. Incontrast, results from our hippocampal model are not consistentwith the idea that recall is a high-threshold process—recall in ourmodel is not all or none, and lures sometimes trigger above-zerorecall. Our claims that lures can trigger some matching recall andthat the number of recalled details varies from item to item areconsistent with the source monitoring framework set forth byMarcia Johnson and her colleagues (see, e.g., Mitchell & Johnson,2000).

Although the recall process in our model is not strictly highthreshold, it is not Gaussian either. If overlap between stimuli isnot too high, our recall process is approximately consistent withdual high-threshold theory in the sense that (for a given experi-ment) there is a level of matching recall that is sometimes ex-ceeded by studied items but not by lures and there is a level ofmismatching recall that is sometimes exceeded by lures but not bystudied items. Furthermore, we assume that participants routinelyset their recall criterion high enough to avoid false recognition ofunrelated lures and that participants set their criterion for recall-to-reject high enough to avoid incorrect rejection of studied items.As such, the model provides some support (conditional on overlapnot being too high) for Yonelinas’s (see Yonelinas et al., 1996)

Figure 29. Effect of medial temporal lobe cortex (MTLC)—specifically,entorhinal cortex input layer (EC in)—damage on forced-choice (FC)recognition performance. This graph plots FC accuracy based on MTLCfamiliarity, hippocampal recall, and a combination of the two signals, as afunction of EC in lesion size. All three accuracy scores (recall alone,familiarity alone, and combined) decline steadily with increasing lesionsize. Hippo � hippocampus.


claim that lures will not be called old based on recall and thatstudied items will not be called new based on recall.

Regarding the independence assumption, as mentioned earlier,Curran and Hintzman (1995) and others have criticized this as-sumption on the grounds that some degree of encoding variabilityis likely to be present in any experiment, and encoding variabilityresults in a positive recall–familiarity correlation. Our results con-firm this latter conclusion, but they also show that interferencereduces the recall–familiarity correlation. As such, it is possible toachieve independence (assuming there is enough interference)even when encoding variability is present.

Overall, although the details of the CLS model’s predictions arenot strictly consistent with the assumptions behind dual-processsignal-detection theory, it is safe to say that the CLS model’spredictions are broadly consistent with these assumptions; accord-ingly, we would expect measurement techniques that rely on theseassumptions to yield meaningful results most of the time. The maincontribution of the CLS model is to establish boundary conditionson the validity of these assumptions—for example, the assumptionthat lures will not be called old based on recall does not hold truewhen there is extensive overlap between stimuli, and the indepen-dence assumption is more likely to hold true in situations whereinterference is high and encoding variability is low than when theopposite is true.

Implications for Theories of How the Hippocampus(Versus the MTLC) Contributes to Recognition

To a first approximation, the CLS model resembles the neuro-psychological theory of episodic memory set forth by Aggletonand Brown (1999; hereafter, the A&B theory)—both theories positthat the hippocampus is essential for recall and that MTLC cansupport judgments of stimulus familiarity on its own. However,this resemblance is only superficial. As discussed in the introduc-tory section, above, simply linking recall and familiarity to hip-pocampus and MTLC, without further specifying how these pro-cesses work, does not allow one to predict when recognition willbe impaired or spared after hippocampal damage. Everythingdepends on how one unpacks the terms recall and familiarity, andthe CLS and A&B theories unpack these terms in completelydifferent ways.

The A&B theory unpacks recall and familiarity in terms of asimple, verbally stated dichotomy whereby recall is necessary forforming new associations but familiarity is sufficient for itemrecognition. In contrast to the A&B theory, the CLS modelgrounds its conception of recall and familiarity in terms of veryspecific ideas about the neurocomputational properties of hip-pocampus and MTLC. Moreover, we have implemented theseideas in a working computational model that can be used to testtheir sufficiency and to generate novel predictions.

According to the CLS model, practically all differences betweencortical and hippocampal contributions to recognition can betraced back to graded differences in how information is repre-sented in these structures. For example, the fact that hippocampalrepresentations are more sparse than cortical representations in ourmodel implies that hippocampal recall is more sensitive to con-junctions than MTLC familiarity, but crucially, both signals shouldshow some sensitivity to conjunctions (see Simulation 5: Associa-tive Recognition and Sensitivity to Conjunctions, above, for more

discussion of this point). This graded approach makes it possiblefor the CLS model to explain dissociations within simple catego-ries such as memory for new associations—for example, the find-ing from Mayes et al. (2001) that hippocampally lesioned patientY.R. sometimes showed intact FC word–word associative recog-nition (presumably, based on MTLC familiarity) despite beingimpaired at cued recall of newly learned paired associates (Mayeset al., 2002; see Vargha-Khadem et al., 1997, for a similar patternof results). As discussed earlier, these dissociations are problem-atic for the A&B theory’s item vs. associative dichotomy.

Lastly, we should mention an important commonality betweenthe CLS model and the A&B theory: Both theories predict sparingof item recognition (with unrelated lures) relative to recall inpatients with complete hippocampal lesions. Although some stud-ies have found this pattern of results in patients with focal hip-pocampal damage (see, e.g., Mayes et al., 2002), it is potentiallyquite problematic for both models that other studies have foundroughly equivalent deficits in recognition (with unrelated lures)and recall following focal hippocampal damage (see, e.g., Manns& Squire, 1999). We do not claim to fully understand why somehippocampally lesioned patients have shown relatively spared itemrecognition but others have not. However, it may be possible toaccount for some of this variability in terms of the idea, set forthby Baxter and Murray (2001b) and explored in Simulation 8:Lesion Effects in the Combined Model, above, that partial hip-pocampal lesions are especially harmful to recognition. This ideais still highly controversial and needs to be tested directly, usingexperiments that carry out careful, parametric manipulations oflesion size (controlling for other factors such as lesion techniqueand task).

Comparison With Abstract Computational Models ofRecognition

Whereas our model incorporates explicit claims about howdifferent brain structures (hippocampus and MTLC) support rec-ognition memory, most computational models of recognitionmemory are abstract in the sense that they do not make specificclaims about how recognition is implemented in the brain. TheREM (i.e., retrieving effectively from memory) model presentedby Shiffrin and Steyvers (1997) represents the state of the art inabstract modeling of recognition memory (see McClelland &Chappell, 1998, for a very similar model). REM carries out aBayesian computation of the likelihood that an ideal observershould say old to an item based on the extent to which that itemmatches (and mismatches) stored memory traces. Our corticalfamiliarity model resembles REM and other abstract models inseveral respects: Like the abstract models, our cortical modelcomputes a scalar that tracks the global match between the testprobe and stored memory traces; furthermore, both our corticalmodel and models such as REM posit that differentiation (Shiffrinet al., 1990) contributes to the null LSE. Thus, our modeling workrelies critically on insights that were developed in the context ofabstract models such as REM.

We can also draw a number of contrasts between the CLS modeland abstract Bayesian models. One major difference is that ourmodel posits that two processes (with distinct operating character-istics) contribute to recognition whereas abstract models attempt toexplain recognition data in terms of a single familiarity process.


Another difference is that—in our model—interference occurs atstudy, when one item reuses weights that are also used by anotheritem, whereas REM posits that memory traces are stored in anoninterfering fashion and that interference arises at test, wheneverthe test item spuriously matches memory traces corresponding toother items.

The question of whether interference occurs at study (or only attest) has a long history in memory research. Other models thatposit structural interference at study have found large and some-times excessive effects of interference on recognition sensitivity.For example, Ratcliff (1990) presented a model consisting of athree-layer feedforward network that learns to reproduce inputpatterns in the output layer; the dependent measure at test is howwell the recalled (output) pattern matches the input. Like ourhippocampal model, the Ratcliff model shows interference effectson recognition sensitivity because recall of discriminative featuresof lures is at floor; as such, any decrease in studied recall neces-sarily pushes the studied and lure recall distributions closer to-gether. The Ratcliff model generally shows more interference thanour hippocampal model because there is more overlap betweenhidden representations in the Ratcliff model than in our hippocam-pal model.

On the basis of these results (and others like them), Murnaneand Shiffrin (1991a) concluded that interference-at-study modelsmay be incapable of explaining the null recognition LSE obtainedby Ratcliff et al. (1990). An important implication of the workpresented here is that interference-at-study models do not alwaysshow excessive effects of interference on recognition sensitivity;our cortical model predicts a null recognition LSE because increas-ing list strength reduces lure familiarity slightly more than studied-item familiarity, so the studied–lure gap in familiarity is preserved.

In this article, we have focused on describing qualitative modelpredictions and the boundary conditions of these predictions.Working at this level, it is clear that there are some fundamentaldifferences in the predictions of the CLS model versus modelssuch as REM. Because studying one item degrades the memorytraces of other items, our model predicts—regardless of parametersettings—that the curve relating interference (e.g., list length or liststrength) to recognition sensitivity will always asymptotically goto zero with increasing interference. In contrast, in REM, it ispossible to completely eliminate the deleterious effects of inter-ference items on performance through differentiation: If interfer-ence items are presented often enough, they can become sostrongly differentiated that the odds of them spuriously matching atest item are effectively zero; whether or not this actually happensdepends on model parameters.

Comparison With Other Neural Network Models ofHippocampal and Cortical Contributions to Recognition

Models of the Hippocampus

The hippocampal component of the CLS model is part of a longtradition of hippocampal modeling (see, e.g., Burgess & O’Keefe,1996; Hasselmo & Wyble, 1997; Levy, 1989; Marr, 1971; Mc-Naughton & Morris, 1987; Moll & Miikkulainen, 1997; Rolls,1989; Touretzky & Redish, 1996; Treves & Rolls, 1994; Wu,Baxter, & Levy, 1996). Although different hippocampal models

may differ slightly in the functions they ascribe to particularhippocampal subcomponents, a remarkable consensus hasemerged regarding how the hippocampus supports episodic mem-ory (i.e., by assigning minimally overlapping CA3 representationsto different episodes, with recurrent connectivity serving to bindtogether the constituent features of those episodes). In the presentmodeling work, we have built on this shared foundation by apply-ing these biologically based computational modeling ideas to arich domain of human memory data (for an application of the samebasic model to animal learning data, see O’Reilly & Rudy, 2001).

The Hasselmo and Wyble (1997) model (hereafter, the H&Wmodel) deserves special consideration because it is the only one ofthe aforementioned hippocampal models that has been used tosimulate patterns of behavioral list-learning data. The architectureof this model is generally similar to the architecture of the CLShippocampal model, except that the H&W model makes a distinc-tion between item and (shared) context information and posits thatitem and context information are kept separate throughout theentire hippocampal processing pathway, except in CA3, whererecurrent connections allow for item–context associations; further-more, in the H&W model, recognition is based on the extent towhich item representations trigger recall of shared contextualinformation associated with the study list. The H&W model pre-dicts that recognition of studied items should be robust to factorsthat degrade hippocampal processing because—insofar as all stud-ied items have the same context vector—the CA3 representation ofshared context information will be very strong and thus easy toactivate. However, the fact that the CA3 context representation iseasy to activate implies that related lures will very frequentlytrigger false alarms in the H&W model (in contrast to the CLSmodel, which predicts low hippocampal false alarms to relatedlures). The H&W model also predicts a null LSE for hippocam-pally driven recognition and a null main effect of item strength onhippocampally driven recognition (in contrast to our model, whichpredicts that both item-strength effects and LSEs should be ob-tained in the hippocampus). Thus, because the H&W model uses adifferent hippocampal recognition measure, as well as separateitem and context representations, it generates recognition predic-tions that are very different from the CLS hippocampal model’spredictions. However, we should emphasize that, if the H&Wmodel used the same recognition measure as our model (match �mismatch) and factored item recall as well as context recall intorecognition decisions, it and the CLS model would likely makevery similar predictions because the two model architectures are sosimilar.

Models of Neocortical Contributions to RecognitionMemory

In recent years, several models besides ours have been devel-oped that address the role of MTLC in familiarity discrimination;see Bogacz and Brown (2003) for a detailed comparison of theproperties of different familiarity-discrimination models. Some ofthese models, like ours, posit that familiarity discrimination incortex arises from Hebbian learning that tunes a population ofunits to respond strongly to the stimulus (see, e.g., Bogacz, Brown,& Giraud-Carrier, 2001; Sohal & Hasselmo, 2000), although thedetails of these models differ from ours (e.g., the Bogacz et al.,


2001, and Sohal & Hasselmo, 2000, models posit that both homo-synaptic Hebbian LTD—which decreases weights if the sendingunit is active but the receiving unit is not—and heterosynapticHebbian LTD—which decreases weights if the receiving unit isactive but the sending unit is not—are important for familiaritydiscrimination, whereas our model incorporates only heterosynap-tic Hebbian LTD). Other models incorporate radically differentmechanisms, for instance, anti-Hebbian learning that reduces con-nection strengths between coactive neurons (Bogacz & Brown,2003).

One difference between the models proposed by Bogacz andBrown (2003) and our model is that—in the Bogacz and Brownmodels—familiarity is computed by a specialized population ofnovelty-detector units that are not directly involved in extractingand representing stimulus features (see Bogacz & Brown, 2003,for a review of empirical evidence that supports this claim). Incontrast, our combined model does not posit the existence ofspecialized novelty-detector units; rather, the layer (EC in) wherethe act win familiarity signal is read out contains the highest (mostrefined) cortical representation of the stimulus, which in turnserves as the input to the hippocampus.

We should note, however, that our model could easily be trans-formed into a model with specialized novelty detectors withoutaltering any of the predictions outlined in the main body of thisarticle. To do this, we could connect the input layer of the com-bined model directly to CA3, CA1, and DG, and we could havethis input layer feed in parallel to a second cortical layer (with noconnections to the hippocampus) where act win is computed. Theunits in this second cortical layer could be labeled specializednovelty detectors insofar as they are not serving any other impor-tant function in the model. This change would not affect thefunctioning of the cortical part of the model in any way.

Having the same layer serve as the input to the novelty-detectionlayer and the hippocampus could, in principle, affect the predic-tions of the combined model, but as a practical matter, none of thecombined model predictions outlined in Simulations 7 and 8,above, would be affected by this change. For example, if corticalunits involved in computing novelty/familiarity were not involvedin passing features to the hippocampus, then it would be possiblein principle to disrupt familiarity for a particular stimulus withoutdisrupting hippocampal recall of that stimulus (by lesioning thenovelty-detector units). However, this is probably not possible inpractice—according to Brown and Xiang (1998), perirhinal neu-rons involved in familiarity discrimination and stimulus represen-tation are topographically interspersed, so lesions large enough toaffect one population of neurons should also affect the other.

A major issue raised by Bogacz and Brown (2003) is whethernetworks such as ours that extract features via Hebbian learninghave adequate capacity to explain people’s ability to discriminatebetween very large numbers of familiar and unfamiliar stimuli(e.g., Standing, 1973, found that people can discriminate betweenstudied and nonstudied pictures after studying a list of thousandsof pictures). Bogacz and Brown argued that—even in a “brain-sized” version of our cortical model—the network’s tendency torepresent shared (prototypical) features at the expense of featuresthat discriminate between items will result in unacceptably poorperformance after studying large numbers of stimuli. Bogacz andBrown pointed out that the anti-Hebbian model that they proposeddoes not have this problem; this model ignores features that are

shared across patterns and, thus, has a much higher capacity. Apossible problem with the anti-Hebbian model is that it may showtoo little interference. More research is needed to assess whetherour network architecture, suitably scaled, can explain findings likethose of Standing (1973) and—if not—how it can be modified toaccommodate this result (without eliminating its ability to explaininterference effects on recognition discrimination).

At this point in time, it is difficult to directly compare ourmodel’s predictions with the predictions of other cortical familiar-ity models because the models have been applied to different datadomains—we have focused on explaining detailed patterns ofbehavioral data, whereas the other models have focused on ex-plaining single-cell recording data in monkeys. Bringing the dif-ferent models to bear on the same data points is an important topicfor future research. Although the CLS model cannot make detailedpredictions about spiking patterns of single neurons, it does makepredictions regarding how firing rates will change as a function offamiliarity. For example, the model predicts that, for a particularstimulus, neurons that show decreased (vs. asymptotically strong)firing in response to repeated presentation of that stimulus shouldbe neurons that initially had a less strong response to the stimulus(and therefore lost the competition to represent the stimulus).

Future Directions

Future research will address limitations of the model that werementioned earlier. In the Sources of Variability section, above, wediscussed how the model incorporates some sources of variabilitythat we plan to remove (sampling variability) and lacks somesources of variability that we plan to add. Increases in computerprocessing speed will make it possible to expand our networks tothe point where sampling variability is negligible, and we willreplace lost sampling variability by adding encoding variabilityand variability in preexperimental presentation frequency to themodel. Including preexperimental variability (by presenting testitems in other contexts a variable number of times prior to the startof the experiment) will allow us to address a range of interestingphenomena, including the so-called frequency mirror effect,whereby hits tend to be higher for low-frequency stimuli than forhigh-frequency stimuli but false alarms tend to be higher forhigh-frequency stimuli than for low-frequency stimuli (see, e.g.,Glanzer, Adams, Iverson, & Kim, 1993); recently, several studieshave obtained evidence suggesting that recall is responsible for thelow-frequency hit-rate advantage and familiarity is responsible forthe high-frequency false-alarm-rate advantage (Joordens & Hock-ley, 2000; Reder et al., 2000; Reder et al. also presented an abstractdual-process model of this finding).

Furthermore, we plan to directly address the question of howparticipants make decisions based on recall and familiarity.Clearly, people are capable of using a variety of different decisionstrategies that can differentially weight the different signals thatemerge from the cortex and hippocampus. One way to address thisissue is to conduct empirical Bayesian analyses to delineate howthe optimal way of making recognition decisions in our modelvaries as a function of situational factors and then compare theresults of these analyses with participants’ actual performance. Aspecific idea that we plan to explore in detail is that participantsdiscount recall of prototype information because prototype recall ismuch less diagnostic than item-specific recall. The frontal lobes


may play an important part in this discounting process—for ex-ample, Curran, Schacter, Norman, and Galluccio (1997) studied afrontal lesioned patient (B.G.) who false-alarmed excessively tononstudied items that were of the same general type as studieditems; one way of explaining this finding is that B.G. has aselective deficit in discounting prototype recall. Thus, the literatureon frontal lesion effects may provide important constraints on howrecognition decision making works by showing how it breaksdown.

Supplementing the model with a more principled theory of howparticipants make recognition decisions will make it possible forus to apply the model to a wider range of recognition phenomena,for example, situations where recall and familiarity are placed inopposition (see, e.g., Jacoby, 1991). We could also begin toaddress the rich literature on how different manipulations affectrecognition ROC curves (see, e.g., Ratcliff et al., 1992; Yonelinas,1994).

Another topic for future research involves improving cross talkbetween the model and neuroimaging data. In principle, we shouldbe able to predict functional magnetic resonance imaging (fMRI)activations during episodic recognition tasks by reading out acti-vation from different subregions of the model; to achieve this goal,we need to build a back end onto the model that relates changes in(simulated) neuronal activity to changes in the hemodynamicresponse that is measured by fMRI. Finally, Curran (2000) hasisolated what appear to be distinct event-related potential (ERP)correlates of recall and familiarity; we should be able to use themodel to predict how these recall and familiarity waveforms willbe affected by different manipulations. Our first attempt alongthese lines was successful; we found that—as predicted by themodel—increasing list strength did not affect how well the ERPfamiliarity correlate discriminated between targets and lures, butlist strength adversely affected how well the ERP recall correlatediscriminated between targets and lures (Norman, Curran, & Tepe,2002).

We also plan to explore other, more biologically plausible waysof reading out familiarity from the cortical network. Althoughact win has the virtue of being easy to compute, it is not imme-diately clear how some other structure in the brain could isolate theactivity of only the winning units (because losing units are stillactive to some small extent and there are many more losing unitsthan winning units). One promising alternative measure is settletime: the time it takes for activity to spread through the network(more concretely, we measured the number of processing cyclesneeded for average activity in MTLC to reach a criterion value of.03). This measure exploits the fact that activity spreads morequickly for familiar than for unfamiliar patterns. The settle timemeasure is more biologically plausible than act win insofar as itrequires only some sensitivity to the average activity of a layer andsome ability to assess how much time elapses between stimulusonset and activity reaching a predetermined criterion. Preliminarysimulation results have shown that settle time yields good d�scores and—like act win—does not show an LSE on d� for ourbasic parameters (20% overlap). Further research is necessary todetermine if the qualitative properties of act win and settle timeare completely identical or if there are manipulations that affectthem differently.

Conclusion

We have provided a comprehensive initial treatment of thedomain of recognition memory using our biologically based neuralnetwork model of the hippocampus and neocortex. This workextends a similarly comprehensive application of the same basicmodel to a range of animal learning phenomena (O’Reilly & Rudy,2001). Thus, we are encouraged by the breadth and depth of datathat can be accounted for within our framework. Future work canbuild on this foundation to address a range of other human andanimal memory phenomena.

References

Aggleton, J. P., & Brown, M. W. (1999). Episodic memory, amnesia, andthe hippocampal–anterior thalamic axis. Behavioral and Brain Sciences,22, 425–490.

Aggleton, J. P., & Shaw, C. (1996). Amnesia and recognition memory: Are-analysis of psychometric data. Neuropsychologia, 34, 51–62.

Atkinson, R. C., & Estes, W. K. (1963). Stimulus sampling theory. In R. D.Luce, R. R. Bush, & E. Galanter (Eds.), Handbook of mathematicalpsychology (pp. 121–268). New York: Wiley.

Bauml, K. (1997). The list-strength effect: Strength-dependent competitionor suppression? Psychonomic Bulletin and Review, 4, 260–264.

Baxter, M. G., & Murray, E. A. (2001a). Effects of hippocampal lesions ondelayed nonmatching-to-sample in monkeys: A reply to Zola and Squire(2001). Hippocampus, 11, 201–203.

Baxter, M. G., & Murray, E. A. (2001b). Opposite relationship of hip-pocampal and rhinal cortex damage to delayed nonmatching-to-sampledeficits in monkeys. Hippocampus, 11, 61–71.

Beason-Held, L. L., Rosene, D. L., Killiany, R. J., & Moss, M. B. (1999).Hippocampal formation lesions produce memory impairments in therhesus monkey. Hippocampus, 9, 562–574.

Bliss, T. V. P., & Collingridge, G. L. (1993, January 7). A synaptic modelof memory: Long-term potentiation in the hippocampus. Nature, 361,31–39.

Bogacz, R., & Brown, M. W. (2003). Comparison of computational modelsof familiarity discrimination in perirhinal cortex. Hippocampus, 13,494–524.

Bogacz, R., Brown, M. W., & Giraud-Carrier, C. (2001). Model of famil-iarity discrimination in the perirhinal cortex. Journal of ComputationalNeuroscience, 10, 5–23.

Brown, M. W., & Xiang, J. Z. (1998). Recognition memory: Neuronalsubstrates of the judgement of prior occurrence. Progress in Neurobi-ology, 55, 149–189.

Burgess, N., & O’Keefe, J. (1996). Neuronal computations underlying thefiring of place cells and their role in navigation. Hippocampus, 6,749–762.

Carpenter, G. A., & Grossberg, S. (1993). Normal and amnesic learning,recognition and memory by a neural model of cortico-hippocampalinteractions. Trends in Neurosciences, 16, 131–137.

Clark, S. E., & Gronlund, S. D. (1996). Global matching models ofrecognition memory: How the models match the data. PsychonomicBulletin & Review, 3, 37–60.

Clark, S. E., Hori, A., & Callan, D. E. (1993). Forced-choice associativerecognition: Implications for global-memory models. Journal of Exper-imental Psychology: Learning, Memory, and Cognition, 19, 871–881.

Cohen, N. J., & Eichenbaum, H. (1993). Memory, amnesia, and thehippocampal system. Cambridge, MA: MIT Press.

Cohen, N. J., Poldrack, R. A., & Eichenbaum, H. (1997). Memory for itemsand memory for relations in the procedural/declarative memory frame-work. In A. R. Mayes & J. J. Downes (Eds.), Theories of organicamnesia (pp. 131–178). Hove, England: Psychology Press.


Curran, T. (2000). Brain potentials of recollection and familiarity. Memory& Cognition, 28, 923.

Curran, T., & Hintzman, D. L. (1995). Violations of the independenceassumption in process dissociation. Journal of Experimental Psychol-ogy: Learning, Memory, and Cognition, 21, 531–547.

Curran, T., Schacter, D. L., Norman, K., & Galluccio, L. (1997). Falserecognition after a right frontal lobe infarction: Memory for general andspecific information. Neuropsychologia, 35, 1035.

Dennis, S., & Humphreys, M. S. (2001). A context noise model of episodicword recognition. Psychological Review, 108, 452–477.

Douglas, R. J., & Martin, K. A. C. (1990). Neocortex. In G. M. Shepherd(Ed.), The synaptic organization of the brain (pp. 389–438). Oxford,England: Oxford University Press.

Eichenbaum, H. (2000). Cortical–hippocampal networks for declarativememory. Nature Reviews Neuroscience, 1, 41–50.

Eichenbaum, H., Otto, T., & Cohen, N. J. (1994). Two functional compo-nents of the hippocampal memory system. Behavioral and Brain Sci-ences, 17, 449–518.

Gillund, G., & Shiffrin, R. M. (1984). A retrieval model for both recog-nition and recall. Psychological Review, 91, 1–67.

Glanzer, M., Adams, J. K., Iverson, G. J., & Kim, K. (1993). The regu-larities of recognition memory. Psychological Review, 100, 546–567.

Grossberg, S. (1976). Adaptive pattern classification and universal recod-ing: I. Parallel development and coding of neural feature detectors.Biological Cybernetics, 23, 121–134.

Grossberg, S. (1986). The adaptive self-organization of serial order inbehavior: Speech, language, and motor control. In E. C. Scwab & H. C.Nusbaum (Eds.), Pattern recognition in humans and machines: Vol. 1:Speech perception (pp. 187–294). New York: Academic Press.

Grossberg, S., & Stone, G. (1986). Neural dynamics of word recognitionand recall: Attentional priming, learning, and resonance. PsychologicalReview, 93, 46–74.

Hasselmo, M. E. (1995). Neuromodulation and cortical function: Modelingthe physiological basis of behavior. Behavioural Brain Research, 67,1–27.

Hasselmo, M. E., Bodelon, C., & Wyble, B. P. (2002). A proposed functionfor hippocampal theta rhythm: Separate phases of encoding and retrievalenhance reversal of prior learning. Neural Computation, 14, 793–818.

Hasselmo, M. E., & Wyble, B. (1997). Free recall and recognition in anetwork model of the hippocampus: Simulating effects of scopolamineon human memory function. Behavioural Brain Research, 89, 1–34.

Hebb, D. O. (1949). The organization of behavior. New York: Wiley.Hinton, G. E., & Plaut, D. C. (1987). Using fast weights to deblur old

memories. In Proceedings of the 9th Annual Conference of the CognitiveScience Society (pp. 177–186). Hillsdale, NJ: Erlbaum.

Hintzman, D. L. (1988). Judgments of frequency and recognition memoryin a multiple-trace memory model. Psychological Review, 95, 528–551.

Hintzman, D. L. (2001). Similarity, global matching, and judgments offrequency. Memory & Cognition, 29, 547–556.

Hintzman, D. L., Curran, T., & Oppy, B. (1992). Effects of similiarity andrepetition on memory: Registration without learning. Journal of Exper-imental Psychology: Learning, Memory, and Cognition, 18, 667–680.

Holdstock, J. S., Mayes, A. R., Roberts, N., Cezayirli, E., Isaac, C. L.,O’Reilly, R. C., & Norman, K. A. (2002). Under what conditions isrecognition spared relative to recall after selective hippocampal damagein humans? Hippocampus, 12, 341–351.

Humphreys, M. S., Bain, J. D., & Pike, R. (1989). Different ways to cue acoherent memory system: A theory for episodic, semantic, and proce-dural tasks. Psychological Review, 96, 208–233.

Jacoby, L. L. (1991). A process dissociation framework: Separating auto-matic from intentional uses of memory. Journal of Memory and Lan-guage, 30, 513–541.

Jacoby, L. L., Yonelinas, A. P., & Jennings, J. M. (1997). The relationbetween conscious and unconscious (automatic) influences: A declara-

tion of independence. In J. D. Cohen & J. W. Schooler (Eds.), Scientificapproaches to consciousness (pp. 13–47). Mahwah, NJ: Erlbaum.

Joordens, S., & Hockley, W. E. (2000). Recollection and familiaritythrough the looking glass: When old does not mirror new. Journal ofExperimental Psychology: Learning, Memory, and Cognition, 26, 1534.

Kahana, M. J., Rizzuto, D., & Schneider, A. (2003). Theoretical correla-tions and measured correlations: Relating recognition and recall in fourdistributed memory models. Manuscript submitted for publication.

Kanerva, P. (1988). Sparse distributed memory. Cambridge, MA: MITPress.

Kohonen, T. (1977). Associative memory: A system theoretical approach.Berlin, Germany: Springer-Verlag.

Kroll, N. E. A., Knight, R. T., Metcalfe, J., Wolf, E. S., & Tulving, E.(1996). Cohesion failure as a source of memory illusions. Journal ofMemory and Language, 35, 176–196.

Levy, W. B. (1989). A computational approach to hippocampal function. InR. D. Hawkins & G. H. Bower (Eds.), Computational models of learningin simple neural systems (pp. 243–304). San Diego, CA: AcademicPress.

Li, L., Miller, E. K., & Desimone, R. (1993). The representation ofstimulus familiarity in anterior inferior temporal cortex. Journal ofNeurophysiology, 69, 1918–1929.

Macmillan, N. A., & Creelman, C. D. (1991). Detection theory: A user’sguide. New York: Cambridge University Press.

Malenka, R. C., & Nicoll, R. A. (1993). NMDA receptor-dependentsynaptic plasticity: Multiple forms and mechanisms. Trends in Neuro-sciences, 16, 521–527.

Mandler, G. (1980). Recognizing: The judgment of previous occurrence.Psychological Review, 87, 252–271.

Manns, J. R., & Squire, L. R. (1999). Impaired recognition memory on theDoors and People test after damage limited to the hippocampal region.Hippocampus, 9, 495–499.

Marr, D. (1971). Simple memory: A theory for archicortex. PhilosophicalTransactions of the Royal Society of London, Series B, 262, 23–81.

Mayes, A. R., Holdstock, J. S., Isaac, C. L., Hunkin, N. M., & Roberts, N.(2002). Relative sparing of item recognition memory in a patient withadult-onset damage limited to the hippocampus. Hippocampus, 12, 325–340.

Mayes, A. R., Isaac, C. L., Downes, J. J., Holdstock, J. S., Hunkin, N. M.,Montaldi, D., et al. (2001). Memory for single items, word pairs, andtemporal order in a patient with selective hippocampal lesions. CognitiveNeuropsychology, 18, 97–123.

McClelland, J. L., & Chappell, M. (1998). Familiarity breeds differentia-tion: A subjective-likelihood approach to the effects of experience inrecognition memory. Psychological Review, 105, 724.

McClelland, J. L., & Goddard, N. H. (1996). Considerations arising froma complementary learning systems perspective on hippocampus andneocortex. Hippocampus, 6, 654–665.

McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Whythere are complementary learning systems in the hippocampus andneocortex: Insights from the successes and failures of connectionistmodels of learning and memory. Psychological Review, 102, 419–457.

McNaughton, B. L., & Morris, R. G. M. (1987). Hippocampal synapticenhancement and information storage within a distributed memory sys-tem. Trends in Neurosciences, 10, 408–415.

Miller, E. K., Li, L., & Desimone, R. (1991, November 29). A neuralmechanism for working and recognition memory in inferior temporalcortex. Science, 254, 1377–1379.

Minai, A. A., & Levy, W. B. (1994). Setting the activity level in sparserandom networks. Neural Computation, 6, 85–99.

Mishkin, M., Suzuki, W., Gadian, D. G., & Vargha-Khadem, F. (1997).Hierarchical organization of cognitive memory. Philosophical Transac-tions of the Royal Society of London, Series B, 352, 1461–1467.


Mishkin, M., Vargha-Khadem, F., & Gadian, D. G. (1998). Amnesia andthe organization of the hippocampal system. Hippocampus, 8, 212–216.

Mitchell, K. J., & Johnson, M. K. (2000). Source monitoring: Attributingmental experiences. In E. Tulving & F. I. M. Craik (Eds.), The Oxfordhandbook of memory (pp. 179–195). New York: Oxford UniversityPress.

Moll, M., & Miikkulainen, R. (1997). Convergence-zone episodic mem-ory: Analysis and simulations. Neural Networks, 10, 1017–1036.

Mumby, D. G., Wood, E. R., Duva, C. A., Kornecook, T. J., Pinel, J. P. J.,& Phillips, A. G. (1996). Ischemia-induced object recognition deficits inrats are attenuated by hippocampal lesions before or soon after ischemia.Behavioral Neuroscience, 110, 266–281.

Murnane, K., & Shiffrin, R. (1991a). Interference and the representation ofevents in memory. Journal of Experimental Psychology: Learning,Memory, and Cognition, 17, 855–874.

Murnane, K., & Shiffrin, R. M. (1991b). Word repetitions in sentencerecognition. Memory & Cognition, 19, 119–130.

Murray, E. A., & Mishkin, M. (1998). Object recognition and locationmemory in monkeys with excitotoxic lesions of the amygdala andhippocampus. Journal of Neuroscience, 18, 6568.

Norman, K. A. (2002). Differential effects of list strength on recollectionand familiarity. Journal of Experimental Psychology: Learning, Mem-ory, and Cognition, 28, 1083–1094.

Norman, K., Curran, T., & Tepe, K. (2002, April). Event-related potentialcorrelates of interference effects on recognition memory. Poster pre-sented at the 9th Annual Meeting of the Cognitive Neuroscience Society,San Francisco.

Nowlan, S. J. (1990). Maximum likelihood competitive learning. In D. S.Touretzky (Ed.), Advances in neural information processing systems(Vol. 2, pp. 574–582). San Mateo, CA: Morgan Kaufmann.

Oja, E. (1982). A simplified neuron model as a principal componentanalyzer. Journal of Mathematical Biology, 15, 267–273.

O’Keefe, J., & Nadel, L. (1978). The hippocampus as a cognitive map.Oxford, England: Oxford University Press.

O’Reilly, R. C. (1996). The Leabra model of neural interactions andlearning in the neocortex. Unpublished doctoral dissertation, CarnegieMellon University.

O’Reilly, R. C. (1998). Six principles for biologically-based computationalmodels of cortical cognition. Trends in Cognitive Sciences, 2, 455–462.

O’Reilly, R. C., & Busby, R. S. (2002). Generalizable relational bindingfrom coarse-coded distributed representations. In T. G. Dietterich, S.Becker, & Z. Ghahramani (Eds.), Advances in neural information pro-cessing systems (Vol. 14, pp. 75–82). Cambridge, MA: MIT Press.

O’Reilly, R. C., & McClelland, J. L. (1994). Hippocampal conjunctiveencoding, storage, and recall: Avoiding a tradeoff. Hippocampus, 4,661–682.

O’Reilly, R. C., & Munakata, Y. (2000). Computational explorations incognitive neuroscience: Understanding the mind by stimulating thebrain. Cambridge, MA: MIT Press.

O’Reilly, R. C., Norman, K. A., & McClelland, J. L. (1998). A hippocam-pal model of recognition memory. In M. I. Jordan, M. J. Kearns, & S. A.Solla (Eds.), Advances in neural information processing systems (Vol.10, pp. 73–79). Cambridge, MA: MIT Press.

O’Reilly, R. C., & Rudy, J. W. (2001). Conjunctive representations inlearning and memory: Principles of cortical and hippocampal function.Psychological Review, 108, 311–345.

Raaijmakers, J. G., & Shiffrin, R. M. (1992). Models for recall andrecognition. Annual Review of Psychology, 43, 205–234.

Ratcliff, R. (1990). Connectionist models of recognition memory: Con-straints imposed by learning and forgetting functions. PsychologicalReview, 97, 285–308.

Ratcliff, R., Clark, S., & Shiffrin, R. (1990). The list-strength effect: I. Dataand discussion. Journal of Experimental Psychology: Learning, Mem-ory, and Cognition, 16, 163–178.

Ratcliff, R., & McKoon, G. (2000). Memory models. In E. Tulving &F. I. M. Craik (Eds.), The Oxford handbook of memory (pp. 571–581).New York: Oxford University Press.

Ratcliff, R., Sheu, C.-F., & Gronlund, S. D. (1992). Testing global memorymodels using ROC curves. Psychological Review, 99, 518–535.

Ratcliff, R., Van Zandt, T., & McKoon, G. (1995). Process dissociation,single-process theories, and recognition memory. Journal of Experimen-tal Psychology: General, 124, 352–374.

Reder, L. M., Nhouyvanisvong, A., Schunn, C. D., Ayers, M. S., Angstadt,P., & Hiraki, K. A. (2000). A mechanistic account of the mirror effectfor word frequency: A computational model of remember–know judg-ments in a continuous recognition paradigm. Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 26, 294–320.

Reed, J. M., Hamann, S. B., Stefanacci, L., & Squire, L. R. (1997). Whenamnesic patients perform well on recognition memory tests. BehavioralNeuroscience, 111, 1163–1170.

Reed, J. M., & Squire, L. R. (1997). Impaired recognition memory inpatients with lesions limited to the hippocampal formation. BehavioralNeuroscience, 111, 667–675.

Rempel-Clower, N. L., Zola, S. M., & Amaral, D. G. (1996). Three casesof enduring memory impairment after bilateral damage limited to thehippocampal formation. Journal of Neuroscience, 16, 5233.

Riches, I. P., Wilson, F. A., & Brown, M. W. (1991). The effects of visualstimulation and memory on neurons of the hippocampal formation andthe neighbouring parahippocampal gyrus and inferior temporal cortex ofthe primate. Journal of Neuroscience, 11, 1763–1779.

Rolls, E. T. (1989). Functions of neuronal networks in the hippocampusand neocortex in memory. In J. H. Byrne & W. O. Berry (Eds.), Neuralmodels of plasticity: Experimental and theoretical approaches (pp.240–265). San Diego, CA: Academic Press.

Rolls, E. T., Baylis, G. C., Hasselmo, M. E., & Nalwa, V. (1989). Theeffect of learning on the face selective responses of neurons in the cortexin the superior temporal sulcus of the monkey. Experimental BrainResearch, 76, 153–164.

Rotello, C. M., & Heit, E. (1999). Two-process models of recognitionmemory: Evidence for recall-to-reject. Journal of Memory and Lan-guage, 40, 432–453.

Rotello, C. M., & Heit, E. (2000). Associative recognition: A case ofrecall-to-reject processing. Memory & Cognition, 28, 907–922.

Rotello, C. M., Macmillan, N. A., & Van Tassel, G. (2000). Recall-to-reject in recognition: Evidence from ROC curves. Journal of Memoryand Language, 43, 67–88.

Rudy, J. W., & O’Reilly, R. C. (2001). Conjunctive representations, thehippocampus, and contextual fear conditioning. Cognitive, Affective, andBehavioral Neuroscience, 1, 66–82.

Rudy, J. W., & Sutherland, R. J. (1989). The hippocampal formation isnecessary for rats to learn and remember configural discriminations.Behavioural Brain Research, 34, 97–109.

Rudy, J. W., & Sutherland, R. W. (1995). Configural association theoryand the hippocampal formation: An appraisal and reconfiguration. Hip-pocampus, 5, 375–389.

Rumelhart, D. E., & Zipser, D. (1986). Feature discovery by competitivelearning. In D. E. Rumelhart, J. L. McClelland, & PDP Research Group(Eds.), Parallel distributed processing: Vol. 1. Foundations (pp. 151–193). Cambridge, MA: MIT Press.

Sherry, D. F., & Schacter, D. L. (1987). The evolution of multiple memorysystems. Psychological Review, 94, 439–454.

Shiffrin, R. M., Huber, D. E., & Marinelli, K. (1995). Effects of categorylength and strength on familiarity in recognition. Journal of Experimen-tal Psychology: Learning, Memory, and Cognition, 21, 267–287.

Shiffrin, R. M., Ratcliff, R., & Clark, S. E. (1990). The list-strength effect:II. Theoretical mechanisms. Journal of Experimental Psychology:Learning, Memory, and Cognition, 16, 179–195.

Shiffrin, R. M., & Steyvers, M. (1997). A model for recognition memory:


REM—Retrieving effectively from memory. Psychonomic Bulletin &Review, 4, 145–166.

Snodgrass, J. G., & Corwin, J. (1988). Pragmatics of measuring recognitionmemory: Applications to dementia and amnesia. Journal of Experimen-tal Psychology: General, 117, 34–50.

Sohal, V. S., & Hasselmo, M. E. (2000). A model for experience-dependent changes in the responses of inferotemporal neurons. Network:Computation in Neural Systems, 11, 169.

Squire, L. R. (1987). Memory and brain. Oxford, England: Oxford Uni-versity Press.

Squire, L. R. (1992a). Declarative and nondeclarative memory: Multiplebrain systems supporting learning and memory. Journal of CognitiveNeuroscience, 4, 232–243.

Squire, L. R. (1992b). Memory and the hippocampus: A synthesis fromfindings with rats, monkeys, and humans. Psychological Review, 99,195–231.

Squire, L. R., Shimamura, A. P., & Amaral, D. G. (1989). Memory and thehippocampus. In J. H. Byrne & W. O. Berry (Eds.), Neural models ofplasticity: Experimental and theoretical approaches (pp. 208–239). SanDiego, CA: Academic Press.

Squire, L. R., & Zola-Morgan, S. (1991, September 20). The medialtemporal memory system. Science, 253, 1380–1386.

Standing, L. (1973). Learning 10,000 pictures. Quarterly Journal of Ex-perimental Psychology, 25, 207–222.

Stark, C. E. L., & Squire, L. R. (2002, April). Memory for associations inpatients with hippocampal damage. Poster session presented at the 9thAnnual Meeting of the Cognitive Neuroscience Society, San Francisco.

Sutherland, R. J., & Rudy, J. W. (1989). Configural association theory: Therole of the hippocampal formation in learning, memory, and amnesia.Psychobiology, 17, 129–144.

Teyler, T. J., & Discenna, P. (1986). The hippocampal memory indexingtheory. Behavioral Neuroscience, 100, 147–154.

Touretzky, D. S., & Redish, A. D. (1996). A theory of rodent navigationbased on interacting representations of space. Hippocampus, 6, 247–270.

Treves, A., & Rolls, E. T. (1994). A computational analysis of the role ofthe hippocampus in memory. Hippocampus, 4, 374–392.

Vargha-Khadem, F., Gadian, D. G., Watkins, K. E., Connelly, A., VanPaesschen, W., & Mishkin, M. (1997, July 18). Differential effects of

early hippocampal pathology on episodic and semantic memory. Sci-ence, 277, 376–380.

Wu, X., Baxter, R. A., & Levy, W. B. (1996). Context codes and the effectof noisy learning on a simplified hippocampal CA3 model. BiologicalCybernetics, 74, 159–165.

Xiang, J. Z., & Brown, M. W. (1998). Differential encoding of novelty,familiarity, and recency in regions of the anterior temporal lobe. Neu-ropharmacology, 37, 657–676.

Yonelinas, A. P. (1994). Receiver-operating characteristics in recognitionmemory: Evidence for a dual-process model. Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 20, 1341–1354.

Yonelinas, A. P. (1997). Recognition memory ROCs for item and asso-ciative information: The contribution of recollection and familiarity.Memory & Cognition, 25, 747–763.

Yonelinas, A. P. (2001). Consciousness, control, and confidence: The 3 Csof recognition memory. Journal of Experimental Psychology: General,130, 361–379.

Yonelinas, A. P. (2002). The nature of recollection and familiarity: Areview of 30 years of research. Journal of Memory and Language, 46,441–517.

Yonelinas, A. P., Dobbins, I., Szymanski, M. D., Dhaliwal, H. S., & King,L. (1996). Signal-detection, threshold, and dual-process models of rec-ognition memory: ROCs and conscious recollection. Consciousness andCognition, 5, 418–441.

Yonelinas, A. P., Hockley, W. E., & Murdock, B. B. (1992). Tests of thelist-strength effect in recognition memory. Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 18, 345–355.

Zola, S. M., & Squire, L. R. (2001). Relationship between magnitude ofdamage to the hippocampus and impaired recognition memory in mon-keys. Hippocampus, 11, 92–98.

Zola, S. M., Squire, L. R., Teng, E., Sefanacci, L., Buffalo, E., & Clark,R. E. (2000). Impaired recognition memory in monkeys after damagelimited to the hippocampal region. Journal of Neuroscience, 20, 451–463.

Zola-Morgan, S., Squire, L. R., & Amaral, D. G. (1986). Human amnesiaand the medial temporal region: Enduring memory impairment follow-ing a bilateral lesion limited to field CA1 of the hippocampus. Journalof Neuroscience, 6, 2950–2967.

(Appendixes follow)


Appendix A

Algorithm Details

This appendix describes the computational details of the algorithm thatwas used in the simulations. The algorithm is identical to the Leabraalgorithm described in O’Reilly and Munakata (2000; see also O’Reilly,1998), except that the error-driven learning component of the Leabraalgorithm was not used here; see Grossberg (1976) for a similar model.Interested readers should refer to O’Reilly and Munakata for more detailsregarding the algorithm and its historical precedents.

Pseudocode

The pseudocode for the algorithm that we used is given here, showingexactly how the pieces of the algorithm described in more detail in thesubsequent sections fit together.

Outer loop: Iterate over events (trials) within an epoch. For each event,settle over cycles of updating:

1. At start of settling, for all units:A. Initialize all state variables (activation, v m, etc.).B. Apply external patterns.

2. During each cycle of settling, for all nonclamped units:A. Compute excitatory netinput (ge(t) or �j, Equation A3).B. Compute kWTA inhibition for each layer based on gi

(EquationA6):i. Sort units into two groups based on gi

: top k and remaining k �1 to n.

ii. Set inhibitory conductance gi between gk and gk � 1

(EquationA5).

C. Compute point-neuron activation combining excitatory input andinhibition (equation A1).

3. Update the weights (based on linear current weight values) for allconnections:A. Compute Hebbian weight changes (Equation A7).B. Increment the weights and apply contrast-enhancement (Equation

A9).

Point-Neuron Activation Function

Leabra uses a point-neuron activation function that models the electro-physiological properties of real neurons while simplifying their geometryto a single point. This function is nearly as simple computationally as thestandard sigmoidal activation function, but the more biologically basedimplementation makes it considerably easier to model inhibitory competi-tion, as described below. Furthermore, using this function enables cogni-tive models to be more easily related to more physiologically detailedsimulations, thereby facilitating bridge-building between biology and cog-nition.

The membrane potential Vm is updated as a function of ionic conduc-tances g with reversal (driving) potentials E as follows:

dVm�t�

dt� � �

c

gc�t�g� c�Ec � Vm�t�� , (A1)

with three channels (c) corresponding to excitatory input (e), leak current(l), and inhibitory input (i). Following electrophysiological convention, theoverall conductance is decomposed into a time-varying component gc(t)computed as a function of the dynamic state of the network and a constantg�c that controls the relative influence of the different conductances. Theequilibrium potential can be written in a simplified form by setting theexcitatory driving potential (Ee) to 1 and the leak and inhibitory drivingpotentials (El and Ei) to 0:

Vm �

geg� e

geg� e � glg� l � gig� i, (A2)

which shows that the neuron is computing a balance between excitationand the opposing forces of leak and inhibition. This equilibrium form of theequation can be understood in terms of a Bayesian decision-making frame-work (O’Reilly & Munakata, 2000).

The excitatory net input/conductance ge(t) or �j is computed as theproportion of open excitatory channels as a function of sending activationstimes the weight values:

� j � ge�t� � �xiwij� �1

n �i

xiwij. (A3)

The inhibitory conductance is computed via the k-winners-take-all(kWTA) function described in the next section, and leak is a constant.

Activation communicated to other cells (yj) is a thresholded () sigmoi-dal function of the membrane potential with gain parameter �:

yj�t� �1

�1 �1

Vm�t� � �� , (A4)

where [x]� is a threshold function that returns 0 if x � 0 and x if x � 0.This sharply thresholded function is convolved with a Gaussian noisekernel ( � .005), which reflects the intrinsic processing noise of biolog-ical neurons. This produces a less discontinuous deterministic functionwith a softer threshold that is better suited for graded learning mechanisms(e.g., gradient descent).

kWTA Inhibition

Leabra uses a kWTA function to achieve sparse distributed representa-tions (cf., Minai & Levy, 1994). Although two different versions arepossible (see O’Reilly & Munakata, 2000, for details), only the simplerform was used in the present simulations. A uniform level of inhibitorycurrent for all units in the layer is computed as follows:

gi � gk�1 � q�gk

� gk�1 �, (A5)

where 0 � q � 1 is a parameter for setting the inhibition between the upperbound of gk

and the lower bound of gk � 1. These boundary inhibition

values are computed as a function of the level of inhibition necessary tokeep a unit right at threshold:

gi �

g*e g� e�Ee � � � glg� l�El � �

� Ei, (A6)

where g*e is the excitatory net input without the bias-weight contribution—

this allows the bias weights to override the kWTA constraint.In the basic version of the kWTA function used here, gk

and gk � 1 are

set to the threshold inhibition value for the kth and k � 1th most excitedunits, respectively. Thus, the inhibition is placed exactly to allow k units tobe above threshold and the remainder below threshold. For this version, theq parameter is almost always .25, allowing the kth unit to be sufficientlyabove the inhibitory threshold. We should emphasize that, when themembrane potential is at threshold, unit activation in the model � .25; assuch, the kWTA algorithm places a firm upper bound on the number ofunits showing activation greater than .25, but it does not set an upper boundon the number of weakly active units (i.e., units showing activationbetween 0 and .25).


Activation dynamics similar to those produced by the kWTA functionhave been shown to result from simulated inhibitory interneurons thatproject both feedforward and feedback inhibition (O’Reilly & Munakata,2000). Thus, although the kWTA function is somewhat biologically im-plausible in its implementation (e.g., requiring global information aboutactivation states and using sorting mechanisms), it provides a computa-tionally effective approximation to biologically plausible inhibitory dy-namics.

Hebbian Learning

The simplest form of Hebbian learning adjusts the weights in proportionto the product of the sending (xi) and receiving (yj) unit activations: �wij �xiyj. The weight vector is dominated by the principal eigenvector of thepairwise correlation matrix of the input, but it also grows without bound.Leabra uses essentially the same learning rule used in competitive learningor mixtures of Gaussians (Grossberg, 1976; Nowlan, 1990; Rumelhart &Zipser, 1986), which can be seen as a variant of the Oja normalization (Oja,1982):

�hebbwij � xiyj � yjwij � yj�xi � wij�. (A7)

Rumelhart and Zipser (1986) and O’Reilly and Munakata (2000) showedthat, when activations are interpreted as probabilities, this equation con-verges on the conditional probability that the sender is active given that thereceiver is active.

To renormalize Hebbian learning for sparse input activations, EquationA7 can be rewritten as follows:

�wij � � yj xi�m � wij� � yj�1 � xi��0 � wij��, (A8)

where an m value of 1 gives Equation A7, whereas a larger value canensure that the weight value between uncorrelated but sparsely active unitsis around .5. Specifically, we set m � .5/�m and �m � .5 � qm (.5 � �),where � is the sending layer’s expected activation level, and qm (calledsavg cor in the simulator) is the extent to which this sending layer’saverage activation is fully corrected for (qm � 1 gives full correction, andqm � 0 yields no correction).

Weight Contrast Enhancement

One limitation of the Hebbian learning algorithm is that the weightslinearly reflect the strength of the conditional probability. This linearity canlimit the network’s ability to focus on only the strongest correlations, whileignoring weaker ones. To remedy this limitation, we introduced a contrast-enhancement function that magnifies the stronger weights and shrinks thesmaller ones in a parametric, continuous fashion. This contrast enhance-ment is achieved by passing the linear weight values computed by thelearning rule through a sigmoidal nonlinearity of the following form:

wij �1

1 � � wij

�1 � wij�� , (A9)

where wij is the contrast-enhanced weight value, and the sigmoidal functionis parameterized by an offset and a gain (standard defaults of 1.25 and6, respectively, used here).

Appendix B

Hippocampal Model Details

This section provides a brief summary of key architectural parameters ofthe hippocampal model. Activity levels, layer size, and projection param-eters were set to mirror the consensus view of the functional architecture ofthe hippocampus described, for example, by Squire, Shimamura, andAmaral (1989).

Table B1 shows the sizes of different hippocampal subregions and theiractivity levels in the model. These activity levels are enforced by settingappropriate k parameters in the Leabra k-winners-take-all (kWTA) inhibi-tion function. As discussed in the main text, activity is much more sparsein dentate gyrus and Region CA3 than in entorhinal cortex (EC).

Table B2 shows the properties of the four modifiable projections in thehippocampal model. For each simulated participant, connection weights inthese projections were set to values randomly sampled from a uniform

distribution with mean and variance (range) as specified in the table. TheScale factor listed in the table shows how influential this projection is,relative to other projections coming into the layer, and % Con ( percentageconnectivity) specifies the percentage of units in the sending layer con-nected to each unit in the receiving layer. Relative to the perforant path, themossy fiber pathway is sparse (i.e., each CA3 neuron receives a muchsmaller number of mossy fiber synapses than perforant path synapses) andstrong (i.e., a given mossy fiber synapse has a much larger impact on CA3unit activation than a given perforant path synapse). The CA3 recurrentsand the Schaffer collaterals projecting from CA3 to CA1 are relativelydiffuse, so that each CA3 neuron and each CA1 neuron receive a largenumber of inputs sampled from the entire CA3 population.

Table B1Sizes of Different Subregions and Their Activity Levelsin the Model

Area Units Activity (%)

EC 240 10.0DG 1,600 1.0CA3 480 4.0CA1 640 10.0

Note. EC � entorhinal cortex; DG � dentate gyrus; CA3 and CA1 �regions of the hippocampus.

Table B2Properties of Modifiable Projections in the Hippocampal Model

Projection Mean Var Scale % Con

EC to DG, CA3 (perforant path) .5 .25 1 25DG to CA3 (mossy fiber) .9 .01 25 4CA3 recurrent .5 .25 1 100CA3 to CA1 (Schaffer) .5 .25 1 100

Note. Mean � mean initial weight strength; Var � variance of the initialweight distribution; Scale � scaling of this projection relative to otherprojections; % Con � percentage connectivity; EC � entorhinal cortex;DG � dentate gyrus; CA3 and CA1 � regions of the hippocampus.


(Appendixes continue)

The connections linking EC in to CA1 and CA1 to EC out are notmodified in the course of the simulated memory experiment. Rather, wepretrain these connections so they form an invertible mapping, whereby theCA1 representation resulting from a given EC in pattern is capable ofrecreating that same pattern on EC out. CA1 is arranged into eight col-umns (consisting of 80 units apiece); each column receives input fromthree slots in EC in and projects back to the corresponding three slots inEC out. See O’Reilly and Rudy (2001) for discussion of why CA1 isstructured in columns.

Lastly, our model incorporates the claim, set forth by Michael Hasselmoand his colleagues (see, e.g., Hasselmo & Wyble, 1997), that the hip-

pocampus has two functional modes: an encoding mode, where CA1activity is primarily driven by EC in, and a retrieval mode, where CA1activity is primarily driven by stored memory traces in CA3; recently,Hasselmo, Bodelon, and Wyble (2002) presented evidence that these twomodes are linked to different phases of the hippocampal theta rhythm.Although we find the theta-rhythm hypothesis to be compelling, we de-cided to implement the two modes in a much simpler way—specifically,we set the scaling factor for the EC in to CA1 projection to a large value(6) at study, and we set the scaling factor to zero at test. This manipulationcaptures the essential difference between the two modes without addingunnecessary complexity to the model.

Appendix C

Basic Parameters

Twenty items at study: 10 target items (which are tested) followed by 10interference items (which are not tested).

Twenty-percent overlap between input patterns (flip 16/24 slots).Fixed high recall criterion, recall � .40.Table C1 shows the other basic parameters, most of which are standard

default parameter values for the Leabra algorithm.For those interested in exploring the model in more detail, it can be

obtained from Kenneth A. Norman’s Computational Memory Laboratoryweb site: http://compmem.princeton.edu.

Received August 13, 2001Revision received July 25, 2002

Accepted August 7, 2002 �

Table C1Basic Parameters for the Hippocampal and Cortical Models

Parameter Value Parameter Value

El 0.15 gl 0.235Ei 0.15 gi 1.0Ee 1.00 ge 1.0Vrest 0.15 0.25� .02 600MTLC � .004 Hippo � .01MTLC savg cor .4 Hippo savg cor 1

Note. MTLC � medial temporal lobe cortex; Hippo � hippocampal;savg cor � correction for sending layer average activation.


Date post:	03-Sep-2020
Category:	Documents
Upload:	others
View:	1 times
Download:	0 times

Modeling Hippocampal and Neocortical Contributions to...

Documents